Artificial Intelligence in the Pathology of Gastric Cancer

Histopathologic diagnosis

The pathologic diagnosis of biopsy and resection specimens of GC is primarily conducted based on the histomorphology of H&E slides. Recently developed DL algorithms enabled the accurate detection of cancer and non-cancer cell types as well as the cancer area and stroma in various types of cancers (Fig. 1).

Fig. 1
Representative image of cancer detection on hematoxylin and eosin slides by the DL model. Well-to-moderately differentiated (A) and poorly differentiated (B) gastric tubular adenocarcinoma and adjacent reactive stroma with inflammatory infiltrates. The DL model can not only recognize cancer epithelium (green area) and peritumoral stroma (blue area) but also differentiate the tumor cells (orange dot) and lymphoplasmacytic infiltration (yellow dot) in (B) well-to-moderately differentiated gastric cancer, (C) poorly differentiated gastric cancer, and (D) poorly differentiated gastric cancer.
DL = deep learning.

• Click for larger image
• Download as PowerPoint slide

Among the various studies regarding AI-based histomorphologic diagnosis of GC, a recent study described the development of a GC detection algorithm in biopsy specimens of daily pathologic practice [21]. A DL algorithm to classify gastric biopsy images into 3 categories (negative for dysplasia, tubular adenoma, or carcinoma) was derived from 2,434 WSIs as training and validation sets, and the performance of the DL model was evaluated by 7,432 test set WSIs. The diagnostic performance of the DL model achieved an area under the receiver operating curve (AUROC) value of 0.9790 for 2-tier classification (negative for dysplasia vs. positive). The sensitivity and specificity for diagnosis were 100% and 97.49%, respectively, when only epithelial tumors were included for analysis. Moreover, the average review time per slide was significantly shorter in the DL-assisted digital image viewer group (18.90 seconds; 95% confidence interval [CI], 17.44–20.36) than in the digital image viewer (35.70 seconds; 95% CI, 33.24–38.15) and conventional microscope (44.97 seconds; 95% CI, 41.43–48.52) groups. The common cause of misinterpretation of tubular adenoma was regenerative glandular atypia and intestinal metaplasia. Inflammatory tissues, including ulcer detritus and granulation tissue, were the most common cause of misinterpretation as carcinoma.

Another study by Song et al. [22] also suggested that the standalone performance of AI is sufficiently high to detect and diagnose GC in the H&E WSI images. They trained the deep CNN with 2,123 pixel-level annotated H&E-stained WSI and validated the algorithm with 3,212 WSIs of gastric tissue slides. The DL model revealed a stable performance with an average AUROC of 0.986 (accuracy, sensitivity, and specificity of 87.3%, 99.6%, and 84.3%, respectively), showing reproducible sensitivity and specificity across independent institutions. Furthermore, the diagnostic accuracy of pathologists who used DL models improved compared with that of pathologists who did not use a DL-assisted diagnosis system for both microscopy and WSI. The DL algorithm also identified 2 initially misdiagnosed cases as negative for malignancy, which were overlooked by pathologists. Conversely, the DL model misinterpreted some cases as false positive malignancies, such as those involving mucus retention extravasation, foamy histiocytes, inflammatory exudates, myofibroblast proliferation, overstained images, or folded sections.

Poorly differentiated adenocarcinoma and signet ring cell carcinoma are often inconspicuous and occasionally misdiagnosed as gastritis by pathologists. Kanavati et al. [23] trained the DL model to classify diffuse type GC from H&E slides using 2 different approaches. First, the model was trained to directly classify diffuse type cancer. Alternatively, the 2-stage approach involved an initial step where the model was used to detect adenocarcinoma, followed by a second stage of further classifying the cases into either diffuse or intestinal type. Both models obtained high AUROCs in the range of 0.95–0.99 (average accuracy, sensitivity, specificity of 91.7%, 93.6%, and 91.7%, respectively) in 5 independent test datasets. The rare false-positive and false-negative predictions stemmed from the misinterpretation of lymphoplasmacytic infiltration in lamina propria as cancer cells and missing degenerative cancer cells in necrotic tissue, respectively.

A recent study by Ba et al. [24] highlighted that DL-assisted interpretation of pathologists could enhance the accuracy and efficiency of GC diagnosis. They conducted a multi-reader study of 16 pathologists using 110 WSIs with or without DL assistance. The AUROC in the DL-assisted pathologist group was significantly higher (0.911 vs. 0.863, P=0.003) than that of pathologists without DL assistance. Additionally, pathologists with DL assistance showed higher sensitivity in GC detection than those without assistance (90.63% vs. 82.75%, P=0.010), whereas no significant difference was observed in specificity between the 2 groups (78.34% vs. 79.90%, P=0.468). Additionally, the use of DL assistance reduced the average review time per WSI in the DL-assisted group compared with that in the group without assistance (22.68 vs. 26.37 seconds, P=0.033).

These findings demonstrated that DL algorithms can detect and differentiate GC cells in H&E slides with sufficiently high performance. Nevertheless, nearly all DL algorithms occasionally revealed misinterpretations that could easily be noticed and corrected by pathologists. Therefore, using DL algorithms as a screening tool or assistance modality for pathologists in the diagnosis of GC specimens can be a plausible approach.

Detection of lymph node metastasis

The regional lymph node metastasis is one of the most important prognostic factors in GC as the tumor, node, and metastasis stage is determined by the number of metastatic lymph nodes [36]. The evaluation of lymph node metastasis under the microscope is laborious for pathologists, especially in the case of micrometastasis (size ranging between 0.2 and 2 mm) [37]. Several studies have demonstrated that DL-based cancer detection on lymph nodes can aid pathologists in their daily practice.

Matsushima et al. [25] trained a DL algorithm with H&E slides, including a total of 51 metastasis-positive nodes and 776 metastasis-negative nodes from 20 patients. The developed DL model achieved an excellent performance with an AUROC of 0.9994. The free-response ROC showed perfect sensitivity with only 3 false positives, wherein histiocytes were detected as cancer cells. Furthermore, the model’s performance was validated on an independent dataset, which yielded similar results with an AUROC of 0.9914.

Another study by Wang et al. [26] that used 6,418 WSIs, including 15,362 lymph nodes with 4,974 metastases, also showed that the DL model’s tumor detection performance in metastatic lymph nodes was comparable with that of experienced pathologists. The accuracy of the DL algorithm was 96.9%, with sensitivity and specificity values of 98.5% and 96.1%, respectively. The sensitivity was influenced by the histological tumor subtype, and misdiagnosis by the DL model mainly occurred in cases of mucinous adenocarcinoma and signet ring cell carcinoma. Moreover, the diagnosis of DL-assisted pathologists showed that 4.6% of cases were initially under-staged by pathologists owing to the neglect of micrometastases. The high tumor-area-to-metastatic lymph node-area ratio (≥0.45) was associated with worse prognosis in GC patients of the same N stage, revealing the value as an independent prognostic factor in both univariate (hazard ratio [HR], 2.05; 95% CI, 1.66–2.54; P<0.001) and multivariate (HR, 1.39; 95% CI, 1.10–1.75; P=0.007) analyses. The results suggest that the DL model can not only provide an accurate histopathologic diagnosis but also a novel prognostic factor of lymph node metastasis in H&E slides, which could be difficult to manually quantify by pathologists.

Hu et al. [27] also utilized 921 WSIs of lymph nodes to develop DL algorithms for both lymph node identification and metastatic cancer detection. The model was highly efficient and accurate in detecting metastatic lymph nodes; the accuracy of lymph node quantification was 97.13%, and the positive and negative predictive values were 93.53% and 97.99%, respectively. As revealed by the results of Wang et al. [26], the false negative rate of the DL model in poorly differentiated adenocarcinoma and signet ring cell carcinoma was higher compared with that of well to moderately differentiated adenocarcinoma. An example of a false positive cancer area was the germinal center of the lymph node, which includes histomorphologically bizarre lymphoid cells owing to a high level of proliferation.

Mitosis counting

Mitosis detection is one of the earliest applications of AI because the mitotic count is included in the histopathologic grading of neoplasms, such as breast cancer and soft tissue sarcoma [38, 39]. The mitotic index is also one of the histological prognostic factors of gastrointestinal stromal tumor (GIST) [40] and neuroendocrine tumors [41]. A higher mitotic index is associated with aggressive tumor behavior, including recurrence and distant metastasis, leading to a poorer prognosis. Studies have shown that AI algorithms could successfully count the mitosis of tumor cells for neuroendocrine tumor, invasive breast carcinoma, and leiomyosarcoma [41, 42, 43]; however, its effectiveness in GIST has not yet been demonstrated. As the morphology of mitotic figures is similar across different tumor types, similar AI models could be applied to GIST. The AI-based mitosis detection algorithm could help pathologists in the accurate counting of mitosis with reduced time and labor.

Programmed death-ligand 1 (PD-L1) scoring

In addition to the image analysis of H&E-stained WSI, AI algorithms have been used for IHC interpretation to reduce possible inter-observer variability of reading [44, 45, 46, 47]. Among the various IHC markers, AI-based interpretation of PD-L1 is important in the histopathologic diagnosis of GC (Fig. 2). PD-1 is an inhibitory checkpoint receptor protein found on immune cells, including cytotoxic T-cells [48]. Some tumor cells express elevated PD-L1 expression on their surface, leading to the inactivation of cytotoxic T-cells and the downregulation of immune responses [49]. PD-L1 protein expression is assessed by IHC for selecting patients eligible for immune checkpoint inhibitor (ICI) treatment [50, 51]. Several studies have demonstrated that digital image analysis of the tumor proportion score (TPS) of PD-L1 is concordant with manual assessment by pathologists in various tumors, although AI-aided PD-L1 TPS assessment in GC has not been attempted yet [46, 52, 53]. A DL algorithm for PD-L1 TPS interpretation in non-small cell lung cancer demonstrated that the DL-assisted evaluation of PD-L1 IHC by pathologists could increase the concordance of readings, especially around the clinically relevant TPS value of 1, leading to a more accurate prediction of the ICI response [44].

Fig. 2
Representative image of DL-based PD-L1 CPS quantification from PD-L1 IHC slides of gastric tubular adenocarcinoma. PD-L1 IHC exhibits heterogeneous immunoreactivity by tumor and mononuclear cells, with (A) CPS <1, (B) CPS 1–49, and (C) CPS ≥50. The DL-model can distinguish PD-L1 positive (purple dot) vs. negative (sky-blue dot) tumor cells and PD-L1 positive (yellow dot) vs. negative (red dot) lymphoplasma cells (D-F).
DL = deep learning; PD-L1 = programmed death-ligand 1; CPS = combined positive score; IHC = immunohistochemistry.

• Click for larger image
• Download as PowerPoint slide

Interpretation of the combined positive score (CPS), which is utilized in GC, is more complex than that of TPS because the numerator not only includes the PD-L1 positive tumor cells but also the mononuclear inflammatory cells in the tumoral stroma. A few AI-based PD-L1 CPS interpretation models in urothelial carcinoma and head and neck squamous cell carcinoma have been reported to show good performance [54, 55]. For GC, an AI-based PD-L1 CPS interpretation model has not yet been developed. However, Kim et al. [28] utilized a digital image analysis model, namely, the Aperio imagescope IHC membrane image analysis algorithm, to obtain the PD-L1 CPS, and compared the results with the interpretation of pathologists using 39 cases of GC with a CPS cutoff of 1. The results of the image analysis model and scoring of pathologists were also compared with the clinical responses of patients to pembrolizumab. Although the model was not AI-based, it showed 84.6% concordance with the interpretation of pathologists. Moreover, in predicting pembrolizumab responses, the PD-L1 CPS obtained by the model and the score interpreted by pathologists did not reveal a statistically significant difference (P=0.1856). The development and validation of AI-based CPS interpretation algorithms in GC will be invaluable for the stratification of patients who will benefit from ICI treatment.

Human epidermal growth factor receptor 2 (HER2) scoring

Another application of AI to the IHC of GC is the quantification of HER2. HER2 is known to be associated with tumor cell proliferation, adhesion, migration, and differentiation [56]. The overexpression or amplification of HER2 is identified in a subset of GC cases, and this is particularly important because of the well-established targeted therapeutics, namely, trastuzumab [57]. Current guidelines recommend IHC and in-situ hybridization for detecting the overexpression or amplification of HER2 [58]. HER2 IHC is scored based on the staining intensity (0, 1+, 2+, and 3+) and the proportion of tumor cells with membranous staining. The HER2 score evaluation on IHC images by AI models has been thoroughly studied for breast cancer [59, 60, 61, 62] but has rarely been investigated in GC. Han et al. [29] developed a DL-based HER2 scoring algorithm for GC using a total of 183 HER2-stained WSIs for training, validation, and testing. The DL model successfully distinguished the staining intensity and proportion of tumor cells and showed effectiveness with an accuracy of 94% for HER2 scoring prediction on WSIs.

Prognostic prediction from H&E images

AI algorithms could not only detect and diagnose cancer cells but also predict the clinical outcome of patients based on the histopathology of various types of cancers [63, 64, 65, 66]. Huang et al. [30] recently developed a DL algorithm for the prognostic prediction of GC using 2,333 H&E images from real-world datasets and The Cancer Genome Atlas (TCGA) program data. The algorithm calculated the risk score based on the most suspicious tiles of WSIs and grouped the GC patients into low- and high-risk subgroups based on the H&E morphology. Notably, the risk-predictive tiles represented features of high-grade malignancies, such as necrosis, vascular or neural invasion, and single-cell invasive morphology. The DL-based prediction of overall survival (OS) in GC patients showed C-index values of 0.728 and 0.671 in the training and internal validation sets, respectively, and the performance was also sufficiently high in the independent external validation set (C-index = 0.657). Moreover, the risk score generated by the DL model was a strong prognostic indicator of GC patients both in univariate (HR, 2.414; P<0.0001) and multivariate (HR, 1.803; P=0.043) analyses. This study demonstrated that the DL model could effectively stratify GC patients based on their risk level, thereby enabling the selection of patients who are more likely to benefit from intensive adjuvant therapy and close follow-ups.

Another study by Wei et al. [31] introduced a DL-based model, called MultiDeepCox-SC, that predicts OS in GC patients by integrating histopathological images with clinical and gene expression data. They trained the model using 10-fold cross-validation with 382 WSIs from the TCGA dataset, and the model calculated the risk score using H&E images, age, gender, tumor grade, tumor stage, and gene expression profiles to predict the patient’s risk of death. The C-index of the model had a prognostic accuracy of 0.660 when relying solely on histopathological images; however, the value significantly increased to 0.744 when clinical and genetic risk factors associated with prognosis were included to calculate the risk score. Subgroup analysis revealed a significant difference in OS between the high- and low-risk score groups among GC patients of the same stage. Moreover, the risk score was an independent predictor of OS in the multivariate analysis on the TCGA dataset (HR, 1.555; P=3.53e-08) and the external test dataset (HR, 2.912; P=9.42e-4).

Prediction of molecular subtypes of GC

The 4 molecular subtypes of gastric adenocarcinoma proposed by TCGA project are as follows: 1) Epstein-Barr (EBV)-infected, 2) microsatellite instability (MSI), 3) genomically stable (GS), and 4) chromosomally unstable tumors (chromosomal instability; CIN) [2]. Subsequent studies have revealed that these molecular subgroups were associated with distinct clinical outcomes and therapeutic responses [67, 68]. The prediction of molecular subtypes of GC from H&E images can provide a cost-effective method for patient stratification for clinicians.

Two DL-based models for predicting molecular subgroups of GC have been introduced to date. Wang et al. [69] developed a DL model using 332 H&E-stained WSIs from the TCGA database that were segmented into 372,505 tiles with molecular subgroup and clinical outcome information. The authors split the entire dataset into 10 subsets and each was trained by the individual DL model; subsequently, an ensemble model was constructed according to the average values of the outputs from the 10 models. The ensemble model showed an enhanced prediction of the 4 molecular subtypes compared with the individual models; it achieved AUROC values of 0.785, 0.668, 0.762, and 0.811 for the prediction of GS, EBV, CIN, and MSI subtypes in tile-level images, and AUROC values of 0.897, 0.764, 0.890, and 0.898 for each subtype at the patient-level. Flinner et al. [70] conducted an independent study to develop a DL model to predict the molecular subtypes of GC using 133 WSIs from the TCGA-stomach adenocarcinoma (STAD) dataset for training and 65 United Kingdom Collaborative (UKC) images for validation. They also utilized an ensemble approach based on majority voting on multiple CNNs to maximize predictability. Compared with laboratory-based molecular testing to determine the molecular subtypes of their own GC cohort collected at UKC, the DL model showed the variable levels of prediction of each molecular subtype. The authors presumed that the intratumoral heterogeneity in GC might have resulted in the low predictability in some cases.

Among the molecular subtypes, the response of GC to ICIs, such as pembrolizumab, is known to be associated with EBV positivity or MSI-high status [71]. Therefore, studies have rigorously been conducted to develop AI algorithms to predict the MSI status and EBV infection directly from H&E histology for their use as beneficial predictive biomarkers.

In a retrospective multicenter cohort study, Muti et al. [32] successfully developed and validated a DL model to detect the EBV and MSI status from histology slides of GC. They obtained tissue samples from ten clinical cohorts of more than 2,500 GC patients from 7 countries (South Korea, Switzerland, Japan, Italy, Germany, the UK, and the USA), and trained the DL model to detect MSI and EBV positivity from digitized H&E slides. First, they trained and validated the DL model within each cohort in a 3-fold cross-validation design. Second, they selected the 5 largest cohorts as pooled training datasets and tested them on the 5 remaining cohorts. In the within-cohort cross-validation, the AUROC values ranged from 0.597 to 0.836 for the detection of MSI, and from 0.819 to 0.897 for the detection of EBV. When trained on a pooled cohort, the AUROC values ranged from 0.723 to 0.863 for the detection of MSI, and from 0.672 to 0.859 for the detection of EBV.

Kather et al. [33] conducted an independent study showing that their DL algorithm can directly predict the MSI status in GC from H&E histology. They utilized histology images of 315 patients from the TCGA-STAD dataset for training, whereas a randomly selected group of 99 patients was used as a test set. As GC in Asian patients has different clinicopathological features compared with that in non-Asian patients, and TCGA-STAD mostly comprises non-Asian patients, an additional cohort of Japanese patients (n=185) was used for external validation. The DL algorithm achieved an AUROC value of 0.81 and 0 in the TCGA-STAD and Japanese cohorts, respectively.

In a study based on the Korean population, Jeong et al. [34] developed a DL-based EBV prediction method from H&E-stained WSIs, using 319 slides from the TCGA dataset and 108 slides from an independent institution. An additional 60 WSIs from other institutions were used for external validation. The EBV prediction model achieved an AUROC value of 0.88 and an area under the precision-recall curve (AUPRC) value of 0.65, both of which were higher than the mean AUROC and AUPRC values of 0.75 and 0.41, respectively, achieved by pathologists. Moreover, the model achieved negative predictive value, sensitivity, specificity, precision, and F1-score of 0.98, 0.86, 0.92, 0.60, and 0.71, respectively.

Another study by Lee et al. [35] demonstrated the automated classification of the MSI status in GC tissue slides using the DL model. They utilized a total of 331 WSIs from the TCGA database, including both frozen and formalin-fixed paraffin-embedded (FFPE) tissues, and conducted 10-fold cross-validation. External validation was conducted using 383 WSIs of gastrectomy specimens of Korean patients. The AUROC values of the model were 0.893 and 0.902 for the frozen and FFPE tissues from TCGA, respectively. The model also achieved good performance with an AUROC value of 0.874 on the external validation cohort.

Prediction of individual genetic alterations

The AI algorithm can serve as a screening tool for the molecular profiling of cancer [20, 72, 73, 74]. Currently, targetable genetic alterations of cancers are mainly identified through next-generation sequencing (NGS); however, insufficient or inadequate sampling of the tumor and the relatively high cost of NGS are practical limitations. Therefore, studies have attempted to develop DL-based models to predict genetic alterations directly from H&E images.

Kather et al. [20] reported that genetic mutations, molecular tumor subtypes, gene expression signatures, and standard pathology biomarkers could be predicted from H&E images by DL algorithms in various types of cancers. For GC, 321 TCGA-STAD WSIs were used for the training and evaluation of DL networks by 3-fold cross-validation. Among all tested tumor types, GC showed the highest absolute number of detectable mutations, and the mean cross-validated AUROC value for the top 8 mutations ranged from 0.66 to 0.78. Key driver mutations of GC, including AMER1, TP53, MTOR, FBXW, and PIK3CA, were successfully detected (P<0.05 after false discovery rate correction). Moreover, gene expression signatures of proliferation, hypermutation, stem cell properties, macrophages, homologous recombination repair deficiency, and wound healing could be detected from histology by the DL algorithm. These results suggest that AI can bridge the gap between genomics and phenomics in GC.

Classifying the tumor immune microenvironment as a biomarker

The distribution of tumor-infiltrating lymphocytes (TILs) in the tumor microenvironment (TME) is a key biomarker in predicting the ICI response in various cancers [75]. Previous studies have demonstrated that the TIL distribution patterns in tumor areas and peritumoral stroma could be classified into 3 distinct immune phenotypes: inflamed (TILs located inside the tumor), excluded (TILs confined to the peritumoral stroma), and desert (few TILs in both tumor nests and peritumoral stroma) [76, 77, 78]. Each phenotype encompassed distinct biological traits, and the inflamed phenotype showed the increased expression of PD-L1 by tumor and immune cells, leading to a superior response to immunotherapy compared with other phenotypes [77, 78]. Park et al. [79] developed an AI-powered WSI analyzer of TIL in the TME to classify immune phenotypes and demonstrated that the spatial analysis of TIL was significantly correlated with tumor response and patient prognosis in advanced non-small cell lung cancer patients. As GC is known to have a heterogenous tumor immune microenvironment, namely, considering the amount of TIL, immune cell subtypes, and their distribution, a similar AI-based approach could be applicable to GC to select patients who are likely to benefit from ICI treatment.