The artificial intelligence revolution in gastric cancer management: clinical applications

Abstract

Nowadays, gastric cancer has become a significant issue in the global cancer burden, and its impact cannot be ignored. The rapid development of artificial intelligence technology is attempting to address this situation, aiming to change the clinical management landscape of gastric cancer fundamentally. In this transformative change, machine learning and deep learning, as two core technologies, play a pivotal role, bringing unprecedented innovations and breakthroughs in the diagnosis, treatment, and prognosis evaluation of gastric cancer. This article comprehensively reviews the latest research status and application of artificial intelligence algorithms in gastric cancer, covering multiple dimensions such as image recognition, pathological analysis, personalized treatment, and prognosis risk assessment. These applications not only significantly improve the sensitivity of gastric cancer risk monitoring, the accuracy of diagnosis, and the precision of survival prognosis but also provide robust data support and a scientific basis for clinical decision-making. The integration of artificial intelligence, from optimizing the diagnosis process and enhancing diagnostic efficiency to promoting the practice of precision medicine, demonstrates its promising prospects for reshaping the treatment model of gastric cancer. Although most of the current AI-based models have not been widely used in clinical practice, with the continuous deepening and expansion of precision medicine, we have reason to believe that a new era of AI-driven gastric cancer care is approaching.

Graphical Abstract

Introduction

Gastric cancer is a high-incidence and multi-malignant neoplastic disease, ranking fifth in incidence and fourth in mortality worldwide [1]. Among them, adenocarcinoma accounts for more than 95% of gastric cancer, and the traditional anatomy and Lauren classification methods can no longer meet the needs of gastric cancer treatment. With the rapid development of molecular biology and informatics, the emergence of molecular feature-based classification methods has opened up a new path for the precision treatment of gastric cancer [2]. There are significant regional differences in the incidence of gastric cancer, with East Asia having the highest rates, followed by Eastern and Central Europe [3]. Although reports over the past decade have shown a decreasing trend in the incidence of gastric cancer in some regions and population groups [4], projections indicate that the global burden of gastric cancer will increase by 62% by 2040, with an estimated 1.8 million new cases and 1.3 million deaths due to gastric cancer each year [5]. It is important to note that Helicobacter pylori infection remains a major risk factor for gastric cancer [6]. Successful eradication of H. pylori infection can reduce the incidence of gastric cancer by 34–53% [7]. Prompt radical surgery is particularly critical for patients with early-stage gastric cancer, with a five-year survival rate of more than 90% [8]. Endoscopic submucosal dissection or endoscopic mucosal resection is currently the preferred standard minimally invasive surgical treatment, followed by laparoscopic and conventional open surgery [9]. However, due to the insidious nature of the early symptoms of gastric cancer, more than 80% of patients are diagnosed at an advanced stage, missing the optimal treatment window [10]. The five-year survival rate for patients with advanced gastric cancer who have developed distant metastases is less than 10% [11]. Chemotherapy, radiotherapy, and targeted therapy can prolong patients’ lives to a certain extent, but the challenges and limitations cannot be ignored. In the 1960s, chemotherapy for advanced gastric cancer began. Single agents such as fluorouracil, cisplatin, and anthracyclines did not work well, and a two- or three-drug combination regimen was often used. The median survival of patients with untreated advanced gastric cancer is only eight months, which can be extended to one year with interventional treatment with standard chemotherapy [12]. Since 2010, targeted therapy has shone in the field of cancer treatment. The ToGA trial, a landmark comparison of trastuzumab plus chemotherapy versus chemotherapy alone in HER2-positive advanced gastric cancer, significantly improved survival (13.8 versus 11.1 months) [13]. However, there is a certain “threshold” for the use of many targeted drugs, such as the efficacy of the anti-HER2 drug trastuzumab, which is highly dependent on HER2-positive status. The HER2-positive rate in gastric cancer in China is only 8.8% [14], meaning that most patients still need to rely on chemotherapy as the main treatment. Until 2020, immunotherapy drugs have brought a revolutionary breakthrough in treating advanced gastric cancer. The world-renowned Phase III clinical study of ORIENT-16 demonstrated that the combination of sintilimab and chemotherapy not only significantly reduced the risk of death but also extended median survival to more than one and a half years (up to 18.4 months in selected patient populations) [15]. However, the widespread use of immunotherapy is limited by its response rate, and many patients with gastric cancer have not yet benefited from it. In this context, the wisdom of traditional Chinese medicine (TCM) of “treating the disease before it happens” enlightens us that we should focus on the early prevention and precise treatment of gastric cancer. However, due to the complex and variable risk factors and pathogenesis of gastric cancer, the current methods for early diagnosis, diagnosis, and clinical management of gastric cancer are overstretched [16]. Therefore, the introduction of innovative technologies has become the key to breaking the game - artificial intelligence is fundamentally changing the management of gastric cancer across the board.

Artificial intelligence (AI) combines the essence of computer science with vast datasets, encompassing subfields such as machine learning and deep learning [17]. Machine learning, a branch of artificial intelligence, utilizes a sophisticated blend of algorithms and statistical techniques to learn from data, uncover patterns, and reveal hidden insights. This grants computers the capacity to learn and mimic human decision-making without the need for cumbersome preset programming [18]. Machine learning is categorized into supervised and unsupervised learning, differing in whether the training data includes label information. Supervised machine learning employs labeled training data to establish relationships between inputs and corresponding labels. In contrast, unsupervised learning does not rely on labeled datasets for training machines, resulting in unsupervised predictions that may be less accurate. Deep learning, an emerging field within machine learning, represents a transition from weak artificial intelligence to general intelligence. By constructing multi-level neural networks, deep learning emulates the intricate learning processes of the human brain, integrating feature extraction and classification. Deep learning has demonstrated exceptional capabilities and promising applications in various domains such as image recognition, natural language processing, and speech recognition [19]. (Figure 1)Despite its significant accomplishments, deep learning encounters challenges like data generalization limitations and interpretability issues [20]. This article aims to comprehensively review the latest research advancements in artificial intelligence for diagnosing, treating, and predicting the prognosis of gastric cancer. It also delves into the current application status and potential constraints to offer insights and guidance for the intelligent evolution of gastric cancer management.

Fig. 1

The relationship between artificial intelligence and machine learning and deep learning

AI-driven surveillance strategies for high-risk populations

Helicobacter pylori infection

Although the eradication strategy of Helicobacter pylori has been effective in reducing the risk of gastric cancer, the risk of gastric cancer is still increased in individuals after successful eradication compared with the general population [7,8,9,10,11,12,13,14,15,16,17,18,19,20,21]. For example, Lin et al. developed the scSE-CatBoost classification system to diagnose H. pylori infection by preprocessing endoscopic images with an accuracy of 0.90, a sensitivity of 1.00, a specificity of 0.81, a positive predictive value of 0.82, a negative predictive value of 1.00, and an area under the curve (AUC) of 0.88 [22]. Nakashima et al. developed a computer-aided diagnostic (CAD) system that divided the H. pylori infection status of patients into three categories: uninfected, infected, and eradicated, with an accuracy of 84.2%, 82.5%, and 79.2%, respectively [23]. Leung et al. evaluated the performance of different machine learning models in predicting the risk of gastric cancer after H. pylori eradication, among which the XGBoost model performed best in predicting cancer development (AUC 0.97), both of which showed great potential in gastric cancer screening [24]. In addition, the ABC method, as an innovative candidate for large-scale screening of gastric cancer, effectively distinguishes between high- and low-risk populations for gastric cancer development by combining the detection of Helicobacter pylori infection with serum pepsinogen levels [25]. Murphy et al. used the Minimal Absolute Contraction and Selection Operator (LASSO) to establish a risk stratification model based on serum data of pepsinogen and antibody responses to 13 Helicobacter pylori antigens, and the results showed a higher AUC (0.74 vs. 0.69) than the ABC method, providing a more accurate tool for gastric cancer risk assessment [26].

Chronic atrophic gastritis with intestinal metaplasia

According to the Correa cascade, the pathological evolution of gastric cancer has the characteristics of multi-stage transformation, which proves that early diagnosis of gastric cancer is extremely important to improve the cure rate and prognosis [27]. In this context, Fang et al. used the semi-supervised deep learning GasMIL algorithm to diagnose and grade pathological images of patients with suspected atrophic gastritis. The algorithm showed the best overall performance in diagnosing intestinal metaplasia (AUC 0.884) and atrophic gastritis (AUC 0.877) in the external test set [28]. Iwaya et al. established the DCNN model and ResNet50 to classify images with and without intestinal metaplasia, with a sensitivity of 97.7% and a specificity of 94.6%, which is helpful for the classification of individual gastric cancer risk in patients with chronic gastritis [29]. Shi et al.used a deep-learning DLDA model to train and test patient pathological images and realized the automatic diagnosis of intestinal metaplasia and dysplasia. The AUC was 0.97 in the internal validation cohort and the macro mean AUC was increased from 0.67 to 0.82 in the independent external validation cohort, demonstrating its strong generalization ability. In addition, Shi et al. further carried out a detailed subclassification of dysplasia and achieved accurate diagnosis across hospitals with the help of domain adaptive optimization technology, which opened up a new way for early screening and risk assessment of gastric cancer [30].

Subepithelial lesions of the gastrointestinal tract

Gastrointestinal stromal tumors and gastrointestinal leiomyomas are common subepithelial lesions, with gastrointestinal stromal tumors being considered potential malignancies [31].To identify these lesions more accurately, Hirai et al.utilized artificial intelligence to classify images and efficiently distinguish five types of lesions, including gastrointestinal stromal tumors, leiomyomas, schwannomas, and ectopic pancreas, achieving an overall accuracy of 86.1%. Specifically, the AI system demonstrated 98.8% sensitivity, 67.6% specificity, and 89.3% accuracy in distinguishing gastrointestinal stromal tumors from non-gastrointestinal stromal tumors, providing robust support for clinical decision-making [32]. Yang et al.developed an artificial intelligence system using images for joint diagnosis, resulting in an increase in the accuracy of endoscopists in diagnosing gastrointestinal stromal tumors or gastrointestinal leiomyomas from 73.8 to 88.8%, showcasing the significant potential of AI in enhancing diagnostic accuracy [33]. Zhang et al. systematically reviewed and analyzed seven studies, revealing that the pooled sensitivity, specificity, and positive and negative likelihood ratios (LR) for differentiating GI stromal tumors from other submucosal tumors under AI-assisted endoscopy were 0.92, 0.82, 4.55, and 0.12, respectively. The total diagnostic odds ratio (DOR) and AUC were 64.70 and 0.950, respectively, strongly indicating the high accuracy and reliability of artificial intelligence in automatically diagnosing gastrointestinal stromal tumors during endoscopy [34].

Artificial intelligence empowers gastric cancer imaging

Artificial intelligence has penetrated endoscopy, histopathology, and a wide range of radiological imaging fields (including CT, MRI, X-ray, etc.), and through the precise analysis of these traditional imaging data, key biomarkers have been successfully screened, thereby achieving both accuracy and high efficiency in the gastric cancer diagnosis process.

Practical exploration of image recognition technology in the diagnosis of gastric cancer

Endoscopic image test

Endoscopy is a crucial tool for detecting early gastric tumors. However, due to the illumination effect under various light sources, there are shadow or reflection issues that impact observation clarity and diagnosis accuracy [35]. To address this limitation, Dong et al.developed an interpretable AI system named ENDOANGEL-ED. This system was used to train, validate, and test patients’ images and videos of focal lesions.In human-machine comparisons, ENDOANGEL-ED outperformed all endoscopists, achieving higher accuracy rates in both internal (81.10% vs. 70.61%) and external video (88.24% vs. 78.49%) evaluations [36]. Gong et al.introduced a deep learning system for classifying four types of lesions (advanced gastric cancer, early gastric cancer, dysplasia, and non-tumor) with an accuracy rate of 89.7%. This system enabled the automatic detection and classification of gastric tumors during real-time endoscopy [37]. Yuan et al.developed an artificial intelligence system that achieved an accuracy rate of 85.7% in diagnosing six gastric lesions under white light endoscopy. This rate was comparable to that of senior endoscopists (85.1%) and higher than that of junior endoscopists(78.8%) [38]. Wu et al.utilized 9151 images to establish a system for DCNN to detect early-stage gastric cancer with an accuracy rate of 92.5%. They also used 24,549 images from various stomach parts to train DCNN to identify blind spots during esophagogastroduodenoscopy, achieving 90% or 65.9% accuracy in tasks where the gastric position is divided into 10 or 26 parts.This performance was comparable to that of experts. This study not only demonstrates the feasibility of AI systems in comprehensive stomach examinations, but also provides robust technical support for enhancing endoscopic training quality in the future [39].

The precision of intelligent endoscopy and expert experience

Traditional diagnostic methods have certain limitations, such as complexity, time consumption, labor intensity, and reliance on physician experience [40]. The integration of artificial intelligence in the medical field is revolutionizing the conventional approach to diagnosis and treatment, aiding clinicians in disease diagnosis, enhancing diagnostic precision, and achieving superior performance [41]. For instance, Zhang et al.developed a deep learning system to differentiate six types of gastric lesions, particularly in identifying early gastric cancer and high-grade intraepithelial neoplasia. The accuracy of the CNN was comparable to that of endoscopists, with higher diagnostic specificity and positive predictive value (PPV) than endoscopists (specificity: 91.2% vs. 86.7%; PPV: 55.4% vs. 41.7%) [42]. Ikenoyama et al.compared CNN with endoscopists and found that CNN’s diagnostic speed was remarkable (45.5 ± 1.8 s), significantly faster than physicians (173 ± 66 min). The sensitivity of CNN (58.4%) was notably higher than that of physicians (31.9%, 26.5% higher), while maintaining high specificity (87.3%) and excellent PPV (26.0%) and negative predictive value (NPV) (96.5%), showcasing the substantial advantages of AI in medical diagnosis [43]. Noda et al.utilized a residual neural network to analyze gastric cancer images in just 7 s, with accuracy, sensitivity, specificity, and other key metrics (83.2%, 76.4%, 92.3%, etc.) surpassing those of endoscopists, particularly in lesion analysis (accuracy 86.1%, sensitivity 82.1%, etc.) [44]. He et al. collected multi-source medical imaging data and developed the ENDOANGEL-ME system, demonstrating outstanding performance, with internal and external test sets and video accuracy exceeding 90%, significantly outperforming the average proficiency of senior endoscopists (70.16% ± 8.78%) [45]. Goto et al. employed an artificial intelligence classifier to differentiate between intramucosal carcinoma and submucosal carcinoma, with accuracy, sensitivity, specificity, and F1 values of 77%, 76%, 78%, and 0.768, respectively.The test values were 72.6%, 53.6%, 91.6%, and 0.662 for endoscopists, and 78.0%, 76.0%, and 80.0% for artificial intelligence and endoscopists, respectively, and 0.776, confirming that AI, in collaboration with physicians, enhances the ability to diagnose the depth of invasion of early gastric cancer [46]. Zhang et al.utilized the AIAG system to diagnose challenging cases with a sensitivity of 79.69% and a specificity of 73.26%, both of which outperformed those of general endoscopists. AI significantly enhanced the specificity of intermediate physicians (59.79% vs. 52.62%), while the performance of experts remained consistent, suggesting that AI had a notable impact on the diagnostic capabilities of intermediate physicians in identifying gastric tumors [47]. Niikura et al. analyzed images from 500 patients, including 100 cases of gastric cancer, to compare the diagnostic accuracy 1:1 between AI and expert endoscopists. The AI group exhibited a higher diagnostic accuracy rate than the experts (100% vs. 94.12%) and a superior image diagnosis rate (99.87% vs. 88.17%) [48]. In conclusion, artificial intelligence is emerging as an indispensable aid in the medical field, and its ongoing advancement and refinement are anticipated to significantly narrow the disparity between traditional diagnostic approaches and modern technology, thereby providing substantial momentum to the precise and efficient progress of the medical sector.

Optimize the precise segmentation algorithm for endoscopic images

Symptoms of gastric cancer vary depending on the stage of progression, but artificial segmentation in images remains a challenging task due to the diversity of mucosal features, irregular margins of lesions, and subtle differences from normal mucosa [49]. Zhang et al.proposed an improved Mask R-CNN (IMR-CNN) model to identify and segment gastric cancer lesions in gastroscopic images by identifying and verifying early gastric cancer images, with precision, recall, accuracy, specificity, and F1-Score values of 92.9%, 95.3%, 93.9%, 92.5%, and 94.1%, respectively [50]. Du et al.proposed a three-branch automatic segmentation framework based on co-spatial attention and channel attention (CSA-CA-TB-ResUnet), which achieved a Jaccard similarity index of 84.54%, a threshold Jaccard index of 81.73%, a Dice similarity coefficient of 91.08%, and a pixel-level accuracy of 91.18%, proving that the correlation information between gastroscopic images can improve the accurate segmentation of early gastric cancer lesions [51]. Sun et al.proposed a new network for gastric lesion segmentation using generative adversarial training, and the experimental results showed that the dice value, accuracy, and recall rate were 86.6%, 91.9%, and 87.3%, respectively, which were significantly better than the existing models. All of the above studies show that artificial intelligence can help provide high-precision, fast, consistent, and real-time automated solutions to improve the performance of endoscopic image segmentation, thereby enhancing the diagnostic efficiency and accuracy of clinicians [52].

Achieve a double reduction in the rate of missed detection and false diagnosis

Due to the delicate nature of gastric tumor lesions and differences in endoscopist skills, clinicians may miss or even misdiagnose during endoscopy [53]. Wu et al.were able to effectively reduce the rate of missed detection of gastric tumors (6.1% vs. 27.3%) in the AI priority group and the conventional priority group without increasing the examination time, while minimizing unnecessary biopsies [54]. Namikawa et al. developed an artificial intelligence diagnostic system to classify gastric cancer and gastric ulcer at the lesion level to reduce the clinical misdiagnosis rate, with an overall accuracy rate of 45.9% (100% for gastric cancer; gastric ulcer 0.8%) and 95.9% (gastric cancer 99.0%; gastric ulcer 93.3%) [55]. With the assistance of artificial intelligence, doctors can use patient information more effectively to improve the accuracy of diagnoses, thereby reducing tragedies caused by misdiagnoses or missed diagnoses.

Innovative application of pathological analysis in the diagnosis of gastric cancer

Endoscopy and pathological biopsy are currently the preferred methods for diagnosis, and pathology is the “gold standard” for tumor diagnosis [56]. However, with the continuous development of science and technology, pathology still faces a series of serious challenges in the modern medical system. Among them, the shortage of pathologists has become an important factor restricting the development of pathology, especially in remote areas and county hospitals with relatively scarce medical resources. At the same time, long-term and high-intensity workloads may lead to a lack of energy among pathologists, which, combined with differences in physicians’ clinical experience and the inevitability of subjective interpretation, increase the uncertainty of diagnostic results [57]. The rise of artificial intelligence (AI) technology has injected unprecedented vitality and transformation into this traditional field. With its powerful data processing capabilities and deep learning algorithms, AI can realize efficient and automated analysis of full-slide digital images (WSI), accurately identify lesion types, and carefully assess the degree of lesions, thereby providing strong auxiliary support for young clinicians [58].

Digital transformation of pathological diagnosis of gastric cancer

With the in-depth penetration and integration of artificial intelligence technology in the medical field, a revolutionary pathological diagnosis paradigm, “digital pathology and intelligent pathology, is gradually emerging, leading the international pathology discipline to a more convenient, intelligent, and efficient development track [57]. In the exploration and practice of “digital pathology, many advanced medical institutions have taken the lead in setting an example and successfully deployed several application cases, demonstrating its huge clinical application potential and value. For example, Hu et al. published a publicly available database of subsize images of gastric histopathology (GasHisSDB) to identify the performance of classifiers. It has the highest accuracy rate of 86.08% and the lowest accuracy rate of 41.12% in traditional machine learning. The best accuracy rate of deep learning is 96.47%, and the lowest accuracy is 86.21% [59]. Song et al.reported that the clinically applicable CNN system developed by the Chinese People’s Liberation Army General Hospital achieved nearly 100% sensitivity and 80.6% average specificity on real-world test datasets after being trained on 2123 pixel-level annotated H&E stained gastric cancer WSI [60]. Ba et al.conducted a fully intersecting multi-reader and multi-case study based on WSI and realized that when deep learning-assisted pathologists explained 110 WSIs, its AUC of 0.911 vs. 0.863) was higher than that of unassisted pathologists, indicating that the application of deep learning improved the accuracy and efficiency of pathologists in the diagnosis of gastric cancer [61]. Tan et al. developed and validated an artificial intelligence radiopathology model with AUCs of 0.953 and 0.851 in the training and validation cohorts, respectively, showing excellent performance in differentiating between stage I-II and stage III gastric cancers [62]. Huang et al.developed a deep-learning model to collect 2333 pathological images from 1037 patients and achieved an accuracy of 0.920 in the external validation set [63]. Veldhuizen et al.constructed a histological subtype classifier for gastric cancer based on deep learning and performed internal validation of fivefold crossover in 731 patients in three cohorts, achieving an average AUC of 0.93 ± 0.07 [64]. Liu et al. demonstrated that their method has the potential for simultaneous gastroscopy and histopathological diagnosis and achieved rapid and highly sensitive imaging and diagnosis of biopsy samples by integrating femtosecond SRS (femto-SRS) and U-Net technology, which won valuable time for early gastric cancer screening and treatment [65].

Pathological diagnosis of lymph node metastasis in gastric cancer

Lymph node metastasis is a significant pathway for the spread of gastric cancer, particularly in advanced stages, where the rate of lymph node metastasis exceeds 70% [66]. Pathologists meticulously examine all anatomically derived lymph node sections under light microscopy to evaluate the status of each lymph node and the total count of lymph nodes in each patient’s section. Despite these efforts, the risk of misdiagnosis remains challenging to eliminate. Due to technical constraints, the actual number of lymph nodes retrieved may fall short of what is necessary for accurate prognosis, leading to inaccurate N staging. To address this limitation, the industry has attempted to introduce the proportion of lymph node metastases as a supplementary indicator of AJCC N staging. However, its constraints cannot be disregarded, and it is challenging to entirely replace the traditional N staging criteria. Therefore, Wang et al. successfully utilized deep learning to predict gastric cancer from resected lymph node WSI and disclose the tumor-to-lymph node metastasis area ratio [67]. Muti et al.developed and validated a deep learning system to directly predict LNM status from a WSI cohort of 1146 patients, achieving accuracy rates of 0.71, 0.69, and 0.65, respectively [68]. Hu et al.segregated 921 lymph node WSIs into a training group and a test group based on a deep-learning algorithm. They utilized 327 unlabeled images for prospective testing to validate the diagnostic system’s performance, achieving an accuracy rate of 97.13% for lymph node quantification.The Xception and DenseNet-121 combined model attained a PPV of 93.53% and an NPV of 97.99%, demonstrating high efficiency and accuracy in lymph node diagnosis to aid pathologists in the initial screening for lymph node metastasis in gastric cancer patients [69]. Zhang et al.devised a 3D enhanced feature pyramid network and conducted experiments on CT image datasets gathered from four medical centers. The results confirmed that the proposed method enhanced the diagnostic accuracy of lymph node metastasis, surpassing existing methods [70]. Lymphovascular invasion is a prognostic factor for gastric cancer, with a higher likelihood of lymph node metastasis correlating with a poorer overall prognosis for the patient [71].Lee et al.introduced a deep learning-based method for detecting lymphovascular infiltration (+), with the ConViT model outperforming AUROC and AURPC using H&E staining WSI in a multi-categorical model (AUROC: 0.9796; AUPRC: 0.9648). This advancement is anticipated to advance precision medicine by saving time and labor in examinations and reducing discrepancies among pathologists [72].

Construct an efficient and low-burden clinicopathological biopsy system

Pathologists meticulously analyze large tissue samples and conduct a thorough pathological examination to identify the type, severity, and extent of diseases. This process involves examining tissue sections under a microscope, diagnosing diseases, assessing lesion characteristics, and collaborating with clinicians to devise treatment plans. Simultaneously, they collaborate closely with surgeons, radiology departments, and other teams to ensure the accuracy of diagnoses and treatments. This role requires high levels of skill and is often performed under tight deadlines, presenting a challenging task for clinical pathologists. To ease the workload of pathologists in diagnosing gastric biopsies, Abe et al.introduced the AL gastric biopsy pathological diagnosis system (AI-G), which utilized the pathological multi-stage semantic segmentation (MSP) method to analyze the distribution of feature values extracted from whole slide imaging (WSI) slides.The accuracy of MSP AI-G (91.0%) surpassed that of the traditional patch AI-G (PB AI-G) (89.8%) in tissue-level validation. Notably, when tested on cohorts from 10 different institutions, MSP AI-G consistently achieved a higher accuracy rate (0.946) compared to PB AI-G (0.861) in tissue-level analysis [73].Park et al.developed a CNN algorithm to automatically classify 2434 WSIs.The diagnostic performance of the two-tier classification, evaluated by AUROC, was 0.9790, distinguishing between negative (NFD) and positive (all cases except NFD). Specifically for epithelial tumors, the sensitivity and specificity were 1.000 and 0.9749, respectively, indicating its potential as a screening aid system for diagnosing gastric biopsy specimens [74].

Artificial intelligence in the development of gastric cancer treatment regimens

Endoscopic submucosal dissection

Endoscopic submucosal dissection is a well-established method for endoscopic resection that can achieve a high overall resection rate of early gastrointestinal (GI) tumors [75]. Undifferentiated early-stage gastric cancer is among the indications for endoscopic submucosal dissection, but the rate of radical resection remains unsatisfactory.Therefore, Bang et al.developed an extreme gradient boosting classifier to predict radical resection of undifferentiated early gastric cancer before endoscopic submucosal dissection based on the morphological and ecological characteristics of the lesion. The internal validation accuracy, precision, recall, and F1 scores were 93.4%, 92.6%, 99.0%, and 95.7%, respectively [76]. Yun et al.created a model to predict non-curative resection by endoscopic submucosal dissection in patients with early-stage gastric cancer. The risk-scoring model demonstrated good discriminant performance in internal validation (AUC 0.851). However, the procedure is technically demanding and carries a high risk of complications. The difficulty of performing endoscopic submucosal dissection is influenced by the lesion’s location [77]. Therefore, Kim et al.developed a gastric simulator that can be used to practice various positions for gastric endoscopic submucosal dissection. Robotic-assisted or traditional endoscopic submucosal dissection is conducted by two trainee endoscopists in challenging or easy locations. In challenging locations, robotic endoscopic submucosal dissection had significantly shorter operative times than conventional dissection (6.2 vs. 10.2 min), mainly due to faster dissection (220.3 vs. 101.9 mm2/min). The blind dissection rate of robotic endoscopic submucosal dissection was significantly lower than that of conventional methods in challenging locations (17.6 vs. 35.2%) [78]. Yang et al.developed a novel flexible-assisted single-arm endoscopic robot (FASTER) that resulted in significantly shorter dissection times than traditional ESD (7 min vs. 13 min), mainly due to faster dissection (148.6 vs. 97.0 mm2/min). The total operative time of FASTER-assisted endoscopic submucosal dissection was shorter than that of conventional endoscopic submucosal dissection, but the difference was not significant (16 vs. 24 min), and all lesions were completely connected without detectable perforation [79].

Precision surgical treatment: key technologies to reshape the medical model and the future of patients

Precision surgical treatment of gastric cancer involves utilizing advanced imaging technology, genetic testing, and minimally invasive surgical techniques to tailor a personalized surgical plan for each patient’s specific condition. By precisely identifying the cancerous site before surgery and utilizing a real-time navigation system during the procedure, the damage to healthy tissues is minimized, thus reducing the risk of postoperative complications. Additionally, by integrating molecular biology methods, the most effective postoperative adjuvant treatment plan is chosen to enhance the cure rate and quality of life for patients. This signifies that gastric cancer treatment has entered a new era of increased efficiency and safety.

The number of lymph node metastases was assessed preoperatively

Accurate preoperative assessment of the number of lymph node metastases is crucial for tailoring treatment for locally advanced gastric cancer. However, the precision of routine preoperative assays is often inadequate [80]. In response to this challenge, Dong et al.developed a deep-learning image nomogram using CT images from 730 patients with locally advanced gastric cancer. The results indicated an AUC of 0.821 for the main cohort, 0.797 for the external validation cohort, and 0.822 for the international validation cohort, demonstrating its reliability and value in predicting the number of lymph node metastases preoperatively [81]. Zhu et al.utilized six machine-learning algorithms to create a prediction model for lymph node metastasis based on clinical characteristics. They observed a 13.63% incidence of lymph node metastasis in patients with locally advanced gastric cancer and noted a non-linear trend and younger age association with lymph node metastasis, offering new insights for developing more precise individualized treatment plans [82].

A preoperative early recurrence risk prediction model

Although significant progress has been made in multidisciplinary approaches to the treatment of patients with resectable locally advanced gastric cancer, overall survival remains low due to the high risk of recurrence.More than half of the recurrences and tumor-related deaths occur within one year after surgery, and the median recurrence-free survival is approximately 10 months [83]. Therefore, preoperative assessment of the risk of early recurrence is of critical significance in clinical practice.Huang et al.used clinical factors and the deep non-enhanced CT features of the psoas major (L3) level to divide 312 gastric cancer patients into a malnutrition group and a normal group to construct a preoperative malnutrition prediction model for gastric cancer, with an AUC of 0.806 and 0.769 and accuracy of 0.734 and 0.766 respectively. This model can provide a new nutritional status assessment and survival assessment tool for gastric cancer patients [84].Zhang et al. recruited patients with advanced gastric cancer from two centers, extracted radiomics features from preoperative diagnostic CT images, applied machine-learning methods to reduce feature size and construct predictive radiomics features, and used multivariate logistic regression analysis to incorporate radiomics features and clinical risk factors into nomograms. With an AUC of 0.831, 0.826, and 0.806, a radiomics nomogram was successfully created to predict the early recurrence detection of patients with preoperative advanced gastric cancer [85].

Intelligent transformation of intraoperative surgery

Gastric cancer surgery has evolved from traditional open surgery to a new era of minimally invasive surgery, in which laparoscopic and robotic surgery have gradually become the mainstream surgical methods.Gong et al.confirmed that robotic distal gastrectomy is feasible and safe, with superior surgical and oncological outcomes compared with laparoscopic distal gastrectomy, and comparable postoperative recovery and postoperative complication outcomes [86]. Li et al.conducted a large-scale, multicenter retrospective cohort study to compare the short- and long-term outcomes of robotic gastrectomy and laparoscopic gastrectomy for gastric cancer. The overall complication rate was lower in the robotic gastrectomy group (12.6% versus 15.2%) than in the laparoscopic gastrectomy group. Additionally, robotic gastrectomy demonstrated advantages such as less blood loss (126.8 mL versus 142.5 mL), more lymph node dissection (32.5 versus 30.7), and better management of the supra pancreatic region (13.3 versus 11.6) [87]. Tokunaga et al. showed that robot-assisted distal gastrectomy and lymph node dissection were safe and feasible in patients with clinical-stage IA gastric cancer [88]. Although robotic surgery is still in the exploratory stage, it has immeasurable room for development, and it is expected to bring benefits to patients by fully grasping and integrating the broad background and favorable environment for the development of 5G networks [89].

Postoperative peritoneal recurrence and metastasis monitoring

Peritoneal recurrence is the primary mode of recurrence after radical surgery for gastric cancer. The median survival time for advanced patients with peritoneal metastasis is only 3 to 6 months, and the 5-year survival rate is less than 2% [90]. An accurate and personalized prediction of peritoneal recurrence is crucial to identify patients who could benefit from intensive therapy.Therefore, Jiang et al.developed a multi-task deep learning model for predicting peritoneal recurrence and disease-free survival simultaneously using preoperative CT images. The model demonstrated consistently high accuracy in predicting peritoneal recurrence in the training cohort (AUC 0.857), the internal validation cohort (0.856), and the external validation cohort (0.843) [91]. Additionally, Chen et al.established a chemomicroscopy-based intelligent cytology-stimulated Raman molecular cytology. This technology enables rapid and precise detection of peritoneal recurrence in just 20 min, with a sensitivity of up to 81.5%, specificity of 84.9%, and an AUC value of 0.85. This technology shows significant potential for the accurate and swift detection of peritoneal recurrence from gastric cancer [92].

Occult peritoneal metastases often occur in patients with advanced gastric cancer, and the currently available diagnostic tools are inadequate [93]. Dong et al.included 554 patients with advanced gastric cancer, confirmed by laparoscopy (122 positive for occult peritoneal metastases and 432 negative). Radiomic features reflecting the phenotypes of the primary tumor (RS1) and peritoneal region (RS2) were constructed from 266 quantitative image features as predictors of peritoneal metastasis (PM). The AUCs of the training group, internal validation group, and two external validation cohorts were 0.958, 0.941, 0.928, and 0.920, respectively, demonstrating extremely high prediction accuracy [94]. Jiang et al.used a deep learning model with 1978 gastric cancer patients to predict CT-based occult peritoneal metastasis. The results showed that the AUC of the PMetNet model was 0.946, with a sensitivity of 75.4% and specificity of 92.9% in the external validation cohort 1, and an AUC of 0.920, a sensitivity of 98.2%, and specificity in the external validation cohort 2, indicating that the PMetNet model can be a reliable non-invasive tool for early identification of patients with occult peritoneal metastases [95]. Sun et al.accurately predicted occult peritoneal metastasis in gastric cancer patients and the benefits of chemotherapy through a non-invasive and interpretable model developed by preoperative CT images. The diagnostic accuracy of occult peritoneal metastasis increased by 10.13–18.86% with the assistance of radiomics features, bringing a new dawn to clinical practice [96].

Neoadjuvant chemotherapy

The emergence of neoadjuvant chemotherapy has significantly improved the prognosis of advanced gastric cancer. Studies such as MAGIC, CLASSIC, and RESOLVE have demonstrated that neoadjuvant chemotherapy offers benefits like reducing the primary tumor size, lowering the tumor stage, increasing the R0 resection rate, decreasing tumor recurrence, and enhancing overall survival [97–98]. Therefore, early noninvasive screening of patients for neoadjuvant chemotherapy is crucial for the personalized treatment of locally advanced gastric cancer [99]. With advancements in imaging technology, CT technology has become more valuable in assessing the effectiveness of neoadjuvant chemotherapy for gastric cancer.Hu et al. identified radioclinical features from pretreatment CT images of 1060 patients with locally advanced gastric cancer. They proposed a deep learning CS model training cohort (TC) and an internal validation cohort (IVC) of 265 patients in five other centers for external validation (EVC). Deep learning CS demonstrated excellent performance in predicting the response of IVC (AUC, 0.86) and EVC (AUC, 0.82) to neoadjuvant chemotherapy, with good calibration in all cohorts [100]. Zhang et al.used deep learning to analyze the CT scans of 633 patients with locally advanced gastric cancer who underwent NAC from three hospitals. The AUC of the validation cohort was 0.808, 0.755, and 0.752, respectively, which were higher than those of the clinical model [101]. Wang et al.included pre-treatment CT images of 323 patients with locally advanced gastric cancer from multiple centers and found that radiomics features had good discrimination performance in predicting treatment response in the training (AUC, 0.736) and external validation (AUC, 0.679) cohorts [102]. Zhong et al.studied patients with locally advanced gastric cancer treated with neoadjuvant chemotherapy and developed clinical models, deep learning models, and nomograms to predict patient response after neoadjuvant chemotherapy. In the validation cohort, the deep learning DRN achieved an AUC of 0.94 and demonstrated adequate discrimination in response to NAC [103]. Cui et al.recruited 719 patients with locally advanced gastric cancer from four hospitals in China. They developed an ensemble deep learning radiomics nomogram (deep learning RN) with AUCs of 0.829, 0.804, and 0.827 in the internal and two external validation cohorts, respectively, showing good calibration in all cohorts (p > 0.05) [104]. Furthermore, the performance of deep learning RN was significantly superior to that of clinical models (p < 0.001), and decision curve analysis confirmed the practical value of deep learning RN in clinical practice. Deep learning RN was also significantly associated with disease-free survival in patients with locally advanced gastric cancer.

Metastasis and management of unresectable cases of advanced gastric cancer

Immunotherapy and immune checkpoint inhibitors

Immunotherapy is the standard treatment for many tumor types [105]. However, only a small percentage of patients achieve clinical benefit. These novel therapies have low objective durable response rates, and some patients may even experience serious adverse effects, making biomarkers that can predict the efficacy and prognosis of immunotherapy all the more important [106]. Wang et al.proposed a multimodal deep-learning radiomics approach to train patients treated with immunotherapy. In the predicted internal and external validation cohorts, the AUC was 0.791 and 0.812, respectively [107]. Ning et al.used a weighted correlation network analysis (WGCNA) algorithm to identify immune subtype-related genes and successfully identified the best prognostic features in the entire cohort using machine learning [108]. Jiang et al.proposed a non-invasive method to accurately predict the tumor microenvironment state of radiological images by combining radiomics and deep learning to analyze a multi-institutional cohort of gastric cancer patients. In patients treated with checkpoint blockade immunotherapy, the model predicts clinical response and further improves predictive accuracy when combined with existing biomarkers [109].

Immune checkpoint inhibitor therapy is widely used but is only effective in a subset of gastric cancers [110]. Currently, personalized immune subtype prediction tools for immune checkpoint inhibitor efficiency in gastric cancer have yet to be developed [111]. In this field, Huang et al.predicted the response of gastric cancer to checkpoint blockade by machine-learning-based personalized immune subtypes, providing “GSClassifier” as a fine computational framework for gene expression profiling (GEP)-based stratification and PAD subtypes, with high accuracy in predicting gastric cancer response to pembrolizumab (anti-PD-1) in advance (AUC 0.833) [112]. Park et al.demonstrated that ACTA2 expression was associated with survival and immune checkpoint inhibitor response in gastric cancer patients in a cohort of 567 patients. Of the 108 patients treated with immune checkpoint inhibitors, 56% of patients with low ACTA2 expression responded to immune checkpoint inhibitors, compared with 25% of patients with high ACTA2 expression [113]. Li et al.included 21 prospective phase I/II gastric cancer studies. The pathological complete response rate (pCR) was 0.21, the major pathological response rate (MPR) was 0.41, and the R0 resection rate was 0.94. Immune checkpoint inhibitors combined with chemoradiotherapy had the highest efficacy, immune checkpoint inhibitors alone had the lowest efficacy, and the efficacy of immune checkpoint inhibitors and chemotherapy ± antiangiogenic was in the middle [114]. Han et al.demonstrated high accuracy and robustness in predicting response to immune checkpoint inhibitors using a pathomics-driven ensemble model that stratified the response to immune checkpoint inhibitors in the training cohort (AUC 0.985), which was further validated in the internal validation cohort (AUC 0.921) as well as the external validation cohort 1 (AUC 0.914) and external validation cohort 2 (0.927) [115]. The emergence of this innovative tool is expected to serve as a novel and valuable tool to promote precision immunotherapy.

Reshaping the tumor microenvironment: evolutionary mechanisms and regulatory strategies from cold to heat

Considering the limitations of PD-L1 as a single biomarker, we need to understand the complex tumor microenvironment to develop better biomarkers and diagnostics to generate effective immune responses to tumors in a targeted manner [116]. “Cold” tumors refer to tumors that are enriched in myeloid-derived suppressor cells (MDSCs) and regulatory suppressor T cells (T-regs), lack immune response activation and TIL infiltration (“immune desert”),have low TMB, poor antigen presentation, and are naturally insensitive to T cell killing. In contrast, thermal tumors refer to tumors with significant T cell and cytotoxic T cell infiltration, inflammatory cytokine-rich tumor microenvironment, T cell inflammatory gene expression profile, high TIL density, and high TMB [117]. Therefore, turning “cold tumors” into “hot tumors” is crucial to improving the response rate and treatment effect of tumor immunotherapy.Therefore, Chen et al.established a multidimensional TIIC profile by considering the density of CD4FoxP3PD-L1, CD8PD-1LAG3, and CD68STING cells, as well as the spatial organization of CD8PD-1LAG3 T cells. TIIC features can predict the response and prognosis of anti-PD-1/PD-L1 immunotherapy in GC patients [116]. Tumor-infiltrating lymphocyte signature can be used as a new pan-cancer predictive biomarker of anti-PD-1/PD-L1 efficacy [118]. By blocking the transforming growth factor-β β (TGF-β) signaling pathway in the tumor microenvironment, the resistance to anti-PD-1/PD-L1 therapy can be significantly improved, and the effect of anti-tumor immunotherapy can be enhanced [119]. Pomponio et al.combined AI-driven digital image analysis and gene expression profiling to quantify the distribution of TILs and characterize the associated TGFβ pathway activity, demonstrating the potential of the association between “cold” tumor immunophenotype and TGFβ signatures as predictive biomarkers [120]. Chen et al.proposed a deep learning model for pathological image prediction of three immune subtypes (immune rejection subtype, immune desert subtype, and immune-inflammatory subtype), with an accuracy of 80.23%,74.45%, and 68.89% for each patch in the training, validation, and test sets, respectively [121]. Sun et al.used four separate cohorts of patients with advanced solid tumors to develop and validate radiomic signatures that predict response to immunotherapy by combining contrast-enhanced CT images and RNA-seq genomic data from tumor biopsies to assess CD8 cell tumor infiltration [122]. Jiang et al.proposed a biologically guided deep learning method that can simultaneously predict tumor immunity and stromal microenvironment states as well as therapeutic outcomes from medical images [123]. While some studies have shown encouraging results, other clinical applications remain challenging.

EBV-associated gastric cancer and microsatellite instability

According to the unique molecular characteristics of gastric cancer, the cancer genome map divides gastric cancer into Epstein-Barr virus infection (EBV), microsatellite instability (MSI), chromosomal instability (CIN), and genome stabilization (GS).Among them, EBV and MSI are the types that may benefit from immunotherapy, while CIN and GS are less likely to respond to immunotherapy [124]. EBVaGC is confirmed by expensive molecular testing (EBV-encoding RNA (EBER) in situ hybridization) [125]. Therefore, Zheng et al.constructed the EBVNet network and fused it with pathologists. The results showed that the AUC cross-validated from the internal was 0.969, the external dataset of multiple institutes was 0.941, and the cancer genome map dataset was 0.895, indicating that the human-machine fusion significantly improved the diagnostic performance of EBVNet and pathologists [126]. Vuong et al.trained a deep learning algorithm to study 137,184 image patches from 16 tissue microarrays (TMAs) (708 tissue cores),24 WSIs, and 286 gastric cancer WSIs. The classifier can classify EBV-GC image plaques of TMAs and WSI with an accuracy of 94.70%,a recall of 0.936,an accuracy of 0.938,an F1 score of 0.937, and a κ coefficient of 0.909 to predict EBV status from gastric cancer WSI, which is expected to reduce clinical costs and tissue waste [127].

Using deep learning to extract biomarkers from pathology sections of large multicenter datasets improves performance.However such datasets are scarce for gastric cancer, and group learning (SL) can overcome this limitation [128]. Therefore, Saldanha et al.used an SL-based classifier to predict an AUC of 0.8092 and EBV of 0.8372 in microsatellite instability (MSI) prediction [129]. Epstein-Barr virus positivity and microsatellite instability (MSI)/mismatch repair deficiency (dMMR) tumors have been reported to be highly sensitive to immune checkpoint inhibitors, but their detection is often costly [130]. Hinata et al.constructed a deep learning model to analyze 408 cases of gastric adenocarcinoma WSI, including 108 cases of EBV,58 cases of MSI/dMMR, and 242 cases of other subtypes, and detected the subgroup (EBV + MSI/dMMR tumors) with high accuracy in the test case with an AUC of 0.947 [131]. Kather et al.showed that deep residual learning can predict microsatellite instability directly from WSI [132]. Chen et al.135 developed and validated a lncRNAs model that uses a machine learning technique, Support Vector Machine (SVM), for automatic microsatellite instability classification. In the training cohort of 94 GC patients, the AUC was 0.976 and 0.950, respectively, providing strong support for the molecular typing and personalized treatment strategy of gastric cancer [133].

HER2-positive gastric cancer

Human epidermal growth factor receptor 2 (HER2) plays a crucial role in the poor prognosis and pathogenesis of various cancers. In the case of advanced HER2-positive gastric cancer, “trastuzumab + chemotherapy” is the primary treatment option [134]. The DESTINY-Gastric01 trial revealed that trastuzumab deruxtecan (T-DXd) led to significant clinical improvements in patients with HER2-positive gastric cancer [135]. He et al.developed and validated a deep learning model using CT images of patients undergoing anti-HER2 targeted therapy. The model, incorporating multifocal and time series data, achieved one-year AUCs of 0.894 and 0.809 for Nomo-LDLM-2 F using time-series medical images and tumor markers, respectively [136]. Chen et al.demonstrated that (18) F-FDG PET/CT scans could potentially aid in predicting HER2 status and guiding treatment decisions for gastric cancer [137].

Model construction of artificial intelligence in the prediction and prognosis of gastric cancer

Deep learning can identify and integrate a variety of factors related to the prognosis of gastric cancer patients, including imaging data, biomarkers, genomic data, etc., to establish accurate prediction models for gastric cancer prognosis and postoperative risk assessment [138].

Imaging diagnosis

By analyzing imaging data, AI can accurately predict the progression, recurrence risk, and treatment effect of gastric cancer patients, to assist doctors in formulating personalized treatment plans and improve the accuracy of prognosis.Arai et al. developed a machine learning model to retrospectively evaluate 1099 patients with chronic gastritis who underwent esophagogastroduodenal and gastric mucosal biopsy sampling, combined with endoscopic and histological findings. They found that 94 patients (8.55%) developed gastric cancer during an average follow-up period of 5.63 years. The researchers were able to generate a personalized cumulative incidence prediction curve for each patient [139]. Yuan et al.found, for the first time, that both tongue images and tongue coating microbiome can be used as tools for gastric cancer diagnosis, with an AUC of 0.89 for a tongue image-based diagnostic model. The AUC (using genus data) of the tongue coating microbiome-based model reached 0.94, and using species data reached 0.95.This was significantly better than the traditional blood biomarkers [140].

Biomarkers

Build predictive models to identify the risk or early lesions of gastric cancer by analyzing specific molecular markers in blood, urine, or other biological specimens. For example, Cai et al.recruited patients who underwent pre-gastroscopy of serum pepsinogen PG I, PG II, Gastrin-17 (G-17), and anti-Helicobacter pylori IgG antibody concentrations in a nationwide multicenter cross-sectional study. The prevalence of gastric cancer observed in derivative cohorts in the low-risk (≤ 11), intermediate-risk (12–16), or high-risk (17–25) groups was 1.2%, 4.4%, and 12.3%, respectively [141]. Chen et al.revealed a diagnostic model for 10 metabolites of gastric cancer by machine learning analysis of patient plasma samples for targeted metabolomics analysis. It has a sensitivity of 0.905 in the external test set, which is superior to traditional methods utilizing cancer protein markers [142]. Mori et al.demonstrated the development of plasma ghrelin levels as a new marker of gastric mucosal atrophy after Helicobacter pylori eradication.The human leukocyte antigen G (HLA-G) molecule is a checkpoint of the immune response, and its overexpression in cancer is associated with immune evasion, metastasis, poor prognosis, and reduced overall survival [143]. Mejía et al.used ELISA tests to assess plasma levels of soluble HLA-G (sHLA-G) in patients with gastric cancer and benign gastric lesions. It was found that sHLA-G concentrations were higher in gastric cancer patients than in patients with benign lesions, plasma sHLA-G levels were higher in women with GC than in men, and there were significant differences in sHLA-G levels between the benign gastric lesions assessed [144].

Genetic analysis

Genomic analysis can provide prognostic and predictive information to guide clinical care.Cheong et al.used machine learning algorithms to identify 32 gene signatures specific to gastric cancer. By employing unsupervised clustering of these gene signatures at the tumor expression level of patients, four molecular subtypes of prognostic survival were identified. A support vector machine with a linear kernel was then constructed to generate a risk score that predicts five-year overall survival, aiding in the clinical care of gastric cancer patients [145]. Zhang et al.constructed an immunogene-related regulatory network and developed two artificial intelligence survival prediction tools.The consistency indices for 1-year, 3-year, and 5-year survival rates were 0.800,0.809,and0.856,respectively, to evaluate the prognostic characteristics of gastric cancer and the disease-free survival of patients [146]. Wei et al. (149) developed a deep learning model, MultiDeepCox-SC, which predicted the overall survival of gastric cancer patients by integrating whole-slide imaging (WSI), clinical data, and gene expression data. The results included a cancer genome map dataset (hazard ratio 1.555, p = 3.53e-08) and an external test set (hazard ratio 2.912, p = 9.42e-4) [147]. Cai et al.utilized machine learning to analyze a gastric cancer gene expression dataset of 1699 patients from five independent cohorts. They developed and validated the prognostic characteristics of gastric cancer based on individualized gene sets and further explored the regulatory mechanisms and therapeutic targets related to survival in gastric cancer [148].

Lymph node metastases

Artificial intelligence is crucial in predicting survival after gastric cancer surgery. It analyzes a large amount of data to accurately assess patients’ prognosis, optimize treatment options, improve survival rates, and reduce the burden on medical resources. Lymph node metastases play a decisive role in choosing treatment regimens for patients with early-stage gastric cancer [149]. However, there is currently no established protocol to predict the risk of lymph node metastasis before and after endoscopic resection [150]. Therefore, Jin et al.developed a deep learning system to retrospectively analyze preoperative CT images of 1699 patients who underwent gastrectomy and lymph node dissection at two medical centers to predict the number of lymph node metastases in locally advanced gastric cancer. Results: The median AUC, sensitivity, and specificity of the 11 node stations in the external validation cohort were 0.876, 0.743, and 0.936 [151]. Lee et al.collected clinicopathological data from gastrectomy patients to construct a GBM model.The results showed that lymph node metastasis was found in 12.6% of the gastrectomy group and 4.3% of the endoscopic submucosal dissection group.The accuracy, sensitivity, specificity, and AUC of group 1 and group 2 were 0.566, 0.922, 0.516, 0.867, 0.810, 0.958, 0.803, and 0.944, respectively [152]. Kim et al.developed a nomogram to predict lymph node metastasis risk status by collecting clinicopathological data from patients undergoing radical resection for early-stage gastric cancer.They compared disease-free and recurrence-free survival between gastrectomy patients and patients in the endoscopic dissection group to avoid unnecessary gastrectomy. For patients with a low risk of lymph node metastasis on the nomogram (≤ 3% of the provisional value in this study), overall survival (P = 0.32),disease-free survival (P = 0.47),and recurrence-free rate (P = 0.09) were similar in the endoscopic dissection and gastrectomy groups [153].

Postoperative survival

Accurate risk prediction of overall survival in gastric cancer patients holds significant prognostic value and aids in categorizing patients into different risk groups to benefit from personalized treatment.Tian et al.established a deep risk network based on whole slide imaging (WSI) characteristics.DeepRisk included 1120 gastric cancer patients, and the results showed a better C-index for overall survival (OS) and disease-free survival (DFS) of 0.71 [154]. Cheong et al.studied a subset of 307 patients with stage II-III gastric cancer (receiving D2 gastrectomy plus adjuvant fluorouracil chemotherapy (n = 193) or surgery alone (n = 114)) and a separate cohort of 625 formalin-fixed paraffin-embedded (FFPE) tumor samples treated from the CLASSIC trial. The 5-year overall survival rates of low-risk, intermediate-risk, and high-risk patients were 83.2%, 74.8%, and 66.0%, respectively. Patients who received adjuvant chemotherapy after surgery showed a significant improvement in five-year OS compared to those who received surgery alone (80% versus 64.5%), while there was no such improvement in five-year overall survival in the non-benefit group (72.5% in patients who received chemotherapy plus surgery versus 72.5% in patients who received surgery alone) [155]. Li et al.developed a support vector machine (SVM) model and included 255 patients who underwent surgical resection.The results showed that the area under the curve (AUC) of 5-year overall survival was 0.773 and the disease-free survival was 0.751 [156]. Wu et al.accurately predicted the postoperative survival rate of gastric cancer patients by establishing a deep learning model. The AUC of the deep learning model at 1 year, 3 years, and 5 years after surgery outperformed other machine learning models and the United States Joint Committee on Cancer (AJCC) staging model [157].

Integration and application of artificial intelligence in endoscopic workflow

Since the symptoms of early gastric cancer are often subtle and the lesions can be relatively insidious, endoscopy plays a crucial role in clinical diagnosis as a direct diagnostic method. The integration of artificial intelligence can assist physicians in accurately identifying lesion areas, thereby reducing the risk of misdiagnosis and missed diagnoses, alleviating the workload of healthcare providers, and enhancing the efficiency of examinations. By optimizing the overall diagnostic and treatment processes associated with endoscopy, AI can facilitate the early detection and intervention of gastric cancer, ultimately improving patient survival rates and quality of life.

AI into the endoscopy workflow begins with clearly defining the objectives and identifying existing challenges. This includes specifying the applications of AI in endoscopy, such as lesion detection, image enhancement, and automated reporting, as well as pinpointing issues in the current workflow, such as diagnostic errors and inefficiencies. Subsequently, a substantial collection of endoscopic images and videos is gathered and annotated to ensure comprehensive coverage of various lesions and normal findings.Experts annotate images by marking lesion areas, identifying types, and providing additional information to train artificial intelligence models. They select appropriate algorithms based on the specific task, such as convolutional neural networks for image recognition, and utilize labeled data to train the model, optimizing its accuracy and generalization capabilities.Verify the model’s performance on an independent dataset to ensure reliability and stability. Next, embed the AI model into the endoscopy system to guarantee compatibility with existing hardware and software. The model should be optimized for real-time image processing, minimizing latency, ensuring efficient processing, and enhancing user interface friendliness. Conduct pilot testing and performance evaluations in a clinical setting to assess the AI’s performance in real-world scenarios, ensuring its accuracy and reliability. Additionally, ensure that the system complies with medical regulations and that the AI system adheres to medical device regulations, such as FDA or CE marking.Comply with data privacy regulations, such as HIPAA and GDPR, to protect patient information. Train doctors to effectively use AI systems, interpret results, and integrate these findings into diagnostic processes. Finally, establish a feedback mechanism and regularly update the model to ensure the system remains advanced and accurate. By following this series of steps, AI can enhance the efficiency and accuracy of endoscopic diagnosis in areas such as lesion detection, image enhancement, and automated reporting. It is essential to pay attention to data quality, algorithmic bias, and ethical considerations to ensure that AI is an aid rather than a replacement for the physician’s role (see Fig. 2 for the overall integration process). Here are some examples:

Image recognition & analysis

Artificial intelligence can analyze images or videos captured by an endoscope using deep learning models to automatically identify potential lesion areas. By training neural networks, AI can learn to recognize abnormal signs in the stomach, such as ulcers, lumps, and polyps, thereby assisting doctors in making more accurate diagnoses [158].

Real-time assisted diagnosis

During endoscopy, artificial intelligence can be combined with a live image stream to provide immediate feedback. For example, the AI system can prompt possible lesions in real time and mark suspicious areas when doctors perform endoscopic operations, helping doctors make decisions more quickly and avoid missed diagnoses [54].

Automated image annotation

Artificial intelligence can automatically label endoscopic images to highlight areas where lesions may be present. This annotation feature not only enhances the diagnostic efficiency of the images but also enables doctors to quickly identify the location and nature of the lesions during subsequent reviews [159].

Data integration and analysis

By integrating artificial intelligence with hospital information systems and electronic medical records, Artificial intelligence can analyze patients’ historical data—such as medical history and family history—to deliver personalized risk assessments. This technology assists doctors in comprehensively evaluating the risk of gastric cancer and offers enhanced decision support during endoscopic procedures [160].

Automate report generation

Artificial intelligence can automatically generate diagnostic reports based on data collected during the endoscopy process. Utilizing speech recognition and natural language generation technologies, Artificial intelligence can provide doctors with draft reports, thereby reducing the time and workload associated with manual report writing [161].

Quality control and training

Artificial intelligence can serve as a valuable quality control tool for monitoring the quality of endoscopic procedures. By analyzing extensive examination data, Artificial intelligence offers feedback on the examination process, operational techniques, and outcomes, thereby assisting physicians in enhancing endoscopic technology. Furthermore, Artificial intelligence can also be utilized as a training resource for medical professionals, enabling novice doctors to refine their skills through simulated diagnostic experiences [162].

Remote diagnostics and collaboration

AI facilitates the remote transmission of endoscopic images and diagnostic results to specialized centers, enabling doctors to perform remote consultations and diagnoses using AI systems. This capability is particularly crucial in remote regions or areas with limited access to expert resources [163].

Fig. 2

Artificial intelligence is integrated into endoscopy workflow

Challenges of AI deployment in different healthcare settings

The deployment of artificial intelligence in various healthcare settings presents a distinct set of challenges. First, the standardization and interoperability of medical data are particularly significant. The wide variety of data formats and encoding methods generated by different healthcare facilities, systems, and devices complicates data sharing and integration. Second, privacy and security are critical considerations in AI applications. Given the substantial amount of sensitive information contained in medical data, it is essential to strictly adhere to relevant data protection regulations, such as the General Data Protection Regulation (GDPR) and the Health Insurance Portability and Accountability Act (HIPAA), when deploying AI to ensure that patient privacy is fully safeguarded. Additionally, the application of AI technology often necessitates seamless integration with existing medical workflows and systems. This requires that AI technology be smoothly incorporated into clinical practice without increasing the burden on medical staff, thereby ensuring the efficiency and operability of the technology in practical applications.

On the other hand, the implementation of artificial intelligence may encounter challenges related to trust and acceptance among clinicians. Doctors and nurses may harbor concerns regarding the reliability and accuracy of AI technology, particularly when it is used in diagnostic and treatment decisions. Clinicians seek to understand how algorithms function, making the interpretability of these algorithms especially crucial. Additionally, fairness is a significant issue, particularly concerning different races, genders, and age groups, as AI algorithms can exhibit biases that result in unequal healthcare outcomes. For healthcare facilities with limited resources, such as rural clinics and hospitals in low-income areas, the feasibility and cost-effectiveness of AI technology require careful consideration. These institutions may lack the necessary infrastructure or funding to implement advanced AI systems. Therefore, the widespread adoption of technology must not only adhere to high-quality standards but also ensure affordability.

The need for artificial intelligence varies significantly across different healthcare settings. Urban hospitals often prioritize high-precision and high-efficiency AI diagnostic tools to address complex medical scenarios and accommodate a large volume of patient needs. In contrast, rural clinics tend to focus on enhancing the accessibility and cost-effectiveness of primary medical services through AI, compensating for limited resources and a shortage of qualified professionals. Additionally, telemedicine systems must ensure real-time and accurate remote diagnosis and treatment to meet the needs of patients in diverse geographical locations. Therefore, the application of AI in healthcare must be tailored to the specific environment and requirements to maximize its practical effectiveness and potential for widespread adoption.

Several measures can be implemented to address these issues. First, promoting the standardization of medical data is essential for establishing a secure data-sharing mechanism that enhances data quality and interoperability. Second, it is crucial to ensure that artificial intelligence complies with regulatory requirements, thereby increasing trust among physicians and patients through improved explainability technologies, such as visual analytics and causal reasoning. Simultaneously, algorithmic bias can be mitigated through fairness testing and diversified data training. Additionally, cloud computing can be leveraged to reduce the costs associated with AI computing, enhance the compatibility of AI with systems like electronic health records, optimize human-machine collaboration, and improve the practicality of AI in clinical settings. Finally, in conjunction with policy guidance and industry collaboration, we will promote the responsible application of AI in healthcare, ensuring that technological innovation and ethical compliance develop in tandem. By implementing these measures, we aim to address the key challenges in the application of artificial intelligence in the medical field, facilitate its implementation and widespread adoption in real-world medical scenarios, and ensure the fairness, reliability, and sustainability of technology applications.

Challenges and prospects of artificial intelligence in gastric cancer research

The intelligent engine of future healthcare focuses on AI-driven clinical decision support systems, especially in the field of gastric cancer diagnosis and treatment. Although data based on ClinicalTrials.gov (https://clinicaltrials.gov/) show that 17 AI trials involving gastric cancer have been completed or are underway, most of the existing AI-based models have not yet been implemented, and their significance in real-world clinical practice is often limited(Tables 1, 2, 3).Currently, artificial intelligence applications in gastric cancer research encounter several challenges, including dataset bias, limited interpretability of AI models, potential false positives and negatives in clinical applications, and ethical concerns regarding patient data privacy, regulatory approvals, and clinician accountability.

Data bias can restrict the applicability of models across diverse populations and may even worsen medical inequities. For instance, if the training data predominantly represents a specific ethnic or gender group, the predictive accuracy of artificial intelligence may decline for other populations. To address this issue, we can utilize diverse datasets that encompass various populations, disease states, and healthcare settings. Additionally, implementing fairness constraint algorithms—such as adjusting weights or resampling the data—can help mitigate bias. It is also essential to incorporate external reviews during the model development phase to ensure that data collection and model training adhere to fairness standards.The “black box” problem of artificial intelligence in medical decision-making can undermine the trust of both medical staff and patients, thereby hindering its clinical application. To address this issue, interpretable AI techniques such as attention mechanisms, visualization tools (e.g., Shapley values), and rule-based models are employed in practical settings to help medical practitioners understand the rationale behind AI decision-making. By integrating AI with traditional medical guidelines, we can ensure that AI recommendations align with doctors’ existing knowledge, thereby enhancing trust. Additionally, improving healthcare workers’ understanding of AI-assisted decision-making through education and training will enable them to utilize the insights provided by AI more effectively. It is important to note that AI can produce false positives (misdiagnoses) or false negatives (missed diagnoses) in disease prediction and diagnosis, which can significantly impact patients’ treatment decisions.We recommend adopting a multi-model fusion strategy to enhance the predictive stability of artificial intelligence and minimize potential misjudgments that may arise from relying on a single model. In high-stakes decisions, such as cancer diagnoses, it is essential to incorporate human expert reviews rather than depending solely on AI. Additionally, through dynamic learning and feedback mechanisms, AI can continuously refine its decision-making processes and improve accuracy in practice. By reducing data bias, enhancing interpretability, and optimizing error rates in clinical applications, AI can serve the healthcare industry more safely and impartially, becoming a reliable assistant to physicians rather than a risk factor.

In healthcare, the responsible adoption of artificial intelligence encompasses a variety of ethical issues, particularly concerning patient data privacy, regulatory approvals, and clinician accountability. Patient data privacy is one of the foremost ethical concerns in the application of AI within the medical field. AI systems typically rely on vast amounts of medical data for training and optimization, including patients’ personal health information, medical records, genetic data, and more. Utilizing this data without patient consent or experiencing a data breach can infringe upon patients’ rights to privacy and may lead to serious identity theft and discrimination. To safeguard patient data privacy, healthcare organizations and AI development companies must adhere to stringent privacy regulations, such as the Health Insurance Portability and Accountability Act (HIPAA) and the General Data Protection Regulation (GDPR). Additionally, it is essential to implement data encryption, anonymization, and other technologies to ensure that sensitive information remains secure during data usage. Furthermore, patients should retain the right to control the use of their data, including the ability to consent to or object to the collection and utilization of their information at any time. Regarding regulatory approvals, the deployment of AI in healthcare must undergo rigorous regulatory scrutiny to ensure its safety, effectiveness, and ethical compliance.Currently, many countries and regions lack a robust regulatory framework for AI medical technology. This gap has allowed some unvalidated AI systems to enter the market, potentially causing harm to patients. The opaque decision-making processes of AI can easily result in a “black box effect, the attribution of responsibility when issues arise. Therefore, regulators must enhance the approval process for medical AI, establish clear laws and regulations, and ensure that AI products undergo rigorous clinical trials and efficacy verification before being introduced to the market. Additionally, regulators should mandate that AI systems are sufficiently transparent, enabling them to clearly articulate their decision-making logic to users and avoid reliance on ambiguous algorithms. Clinicians continue to play a crucial role in AI-assisted medical decision-making processes. While AI can provide data analysis and diagnostic recommendations, the final decision should always rest with an experienced clinician.As a result, clinicians must assume ultimate responsibility for their diagnostic results and treatment decisions when utilizing artificial intelligence. Clinician accountability is crucial to ensuring the quality of care and patient safety. If an error in an AI system leads to a medical mistake, it must be clear whether the physician is adhering to the correct procedures and relying on the AI’s results to make informed decisions. To enhance the effectiveness of accountability, physicians should receive ongoing training regarding the capabilities and limitations of AI systems. Additionally, healthcare facilities should implement robust vetting mechanisms to ensure that physicians remain at the forefront of the AI decision-making process.

To address these challenges, we must expedite the development of a medical data-sharing platform and promote data integration and transparency to enhance the applicability of artificial intelligence in the healthcare sector. Simultaneously, it is essential to strengthen the explainability and transparency of algorithms to bolster their credibility in clinical diagnosis and treatment. Furthermore, establishing a robust legal, regulatory, and ethical framework is crucial for enhancing the security and privacy protection of patient data, ensuring the compliant use of AI. Looking ahead, artificial intelligence is poised to continue guiding the evolution of gastric cancer diagnosis and treatment toward greater efficiency and accuracy, ultimately contributing to the reduction or even eradication of gastric cancer as a global health threat.

Table 1 Artificial intelligence gastric cancer diagnosis system

Table 2 Artificial intelligence gastric cancer diagnosis model

Table 3 Artificial intelligence gastric cancer prediction model