Deep learning-based subtyping of gastric cancer histology predicts clinical outcome: a multi-institutional retrospective study

The development and application of DL, Machine Learning (ML) and Artificial Intelligence-based methods and techniques has seen an exponential rise within the field of Computational Pathology over the past decade [1, 2]. In particular, computer vision-based DL techniques using routine H&E-stained tissue sections have opened potential avenues for clinically relevant DL-based assessment of H&E-stained specimens [3‐6].

GC is among the most prevalent cancers worldwide [7], and the vast majority of GC are adenocarcinomas. Since its introduction in 1965, the histopathological classification system of Laurén (hereinafter referred to as Laurén) is one of the most widely used histological classifications for adenocarcinomas of the stomach in the West [8]. It identifies two main histological subtypes: intestinal and diffuse types. These subtypes are genetically distinct and associated with different clinical outcomes. In the advanced disease stage, diffuse type GC are typically considered to be more aggressive resulting in poorer patient prognosis compared with intestinal type GC [9]. Research has been performed investigating the utility of Laurén for patient management decisions and several clinical trials are currently ongoing randomizing patients to different treatment options based on Laurén [9‐14].

GC is well known for its high degree of histological inter- and intratumoral heterogeneity which is most likely the reason for high rates of intra- and interobserver variability of all known GC histological classification systems including Laurén [15, 16]. The inter- and intratumoral heterogeneity of GC and resulting challenges in accurate and reproducible histological classification [17, 18] may render prediction of patient outcomes difficult. The 4th edition of the WHO classification defined for the first time a ‘mixed type’ GC as presence of a poorly cohesive component (diffuse type according to Laurén) in combination with another histological subtype irrespective of relative amounts of each component [19]. This may lead to a further increase in interobserver variation of GC histology classification. Furthermore, the variation in the extent of tumor tissue sampling may lead to potentially ‘misclassified’ cases.

GC incidence is highest in Asian countries. In addition, GC patients from these regions often have higher rates of diffuse type GC and as such many Asian medical societies have developed their own histological classification systems which is more similar to the WHO classification system than to Laurén [20]. That said, conversion tables between classifications have been published [21, 22].

To the best of our knowledge, while intestinal and diffuse type GC has been attempted to be predicted genetically [11], DL-based classification of GC into diffuse and intestinal type has not previously been attempted systematically using large cohorts with clinical follow-up data. We have previously used DL to predict diffuse and intestinal type GC, but these results, while encouraging, were limited to a technical benchmark study [23].

We hypothesized that DL has the potential to reduce the above-mentioned inter- and intra-observer variation in classifying histological subtypes in GC and thus may improve accuracy and reproducibility of Laurén through its usage as a classification aid for pathologists.

The primary objective of the present investigation was to (1) establish a DL-based model to classify GC as diffuse or intestinal type and to (2) test the model performance on digital H&E-stained tissue sections from GC resection samples from European and Asian patients without outlining tumor regions. The secondary objective was to determine the hypothetical utility of such a model when used alongside pathologist classifications.

All specimens were obtained from resection specimens, however, tumor cell morphology changes from the tumor surface to the center and invasion front. Thus, it remains to be shown whether the DL-based classification can be applied to endoscopic tumor biopsies as treatment decisions for GC patients with resectable disease have to be made at the time of the diagnostic biopsy. Additionally, the highly heterogeneous nature of GC means that single tissue sections, as was used in the present study, may not be fully representative of the whole tumor. We also recognize that future work could benefit from investigating strategies for selecting the most representative sections of the tumor, taking into consideration the challenges posed by intratumoral heterogeneity in GC classification. Finally, we recognize the concern of overfitting in the TCGA cohort and acknowledge that some degree of overfitting is inevitable. We nonetheless believe that this study demonstrates the potential of DL-based histology typing as an aid in subtyping, serving as a strong foundation for further investigation.

Laurén is a commonly used method for histologic subtyping in GC. Nevertheless, this classification is subjective with high interobserver variation, and there remains a lack of consensus regarding its value as a prognostic tool. To the best of our knowledge, this is the first study to attempt Laurén prediction in GC using current state-of-the-art DL techniques in a systematic manner. Our results show that a model trained on TCGA GC WSIs using pathologist’s Laurén as ground truth achieves an excellent performance when internally cross-validated with an AUROC of 93%. Furthermore, we show that despite imperfect external validation performance, the DL-based histological subtype was able to stratify patients by survival whereas the pathologist-based Laurén (ground truth) did not. These findings seem to suggest that a DL-based histology subtype model trained on the pathologist’s Laurén classification and not trained on the survival data is able to predict patient survival. Up to 54% of GC classified as intestinal type by the DL model were originally classified as diffuse type by the pathologist. This relatively large number of reclassified cases using the DL-based classification could explain why the pathologist-based Laurén subtypes did not stratify patients by survival. This could reflect the known difficulties pathologists have to decide whether a poorly differentiated cancer should be classified as intestinal type or diffuse type. Further detailed quantitative analyses of the misclassified cases are needed to fully understand the prognostic relationship.

In our study, we observed a discrepancy between the high accuracy of the DL model in the TCGA dataset and its performance in the external datasets. However, it is important to emphasize that our primary objective was to assess the potential prognostic utility of the DL-based classifier for GC histology subtyping, rather than achieving perfect alignment with the Laurén classification. One has to realize that a DL-based model attempting to predict histological phenotype in GC will not fully reproduce the human decision-making process, e.g. the way pathologists establish the histological subtype according to the Lauren classification for a particular case. Pathologists may have a more nuanced view, evaluating the histological features on an intestinal-diffuse continuum to reach a final decision about the histological subtype. This strongly differs from the strictly dichotomous nature of the DL-based classifier based on the majority of patches classified as a certain subtype. Thus, because of the known heterogeneity of GC and the above explained majority call by the DL model, it is expected that during post-hoc evaluation of the DL-based classification by an expert pathologist, some of the DL model classification were found to be inaccurate. Nonetheless, survival stratification of patients appeared to be superior when using DL-based histological subtypes. This could indicate that the model is focused on a subset of histological features which happen to be prognostically relevant which are currently not recognized as such by the pathologist. It would be of clinical value to further scrutinize these histological slides, potentially using a different set of image analysis tools, to identify these prognostically relevant features so pathologists can learn to understand and recognize these themselves in the future. Alternatively, one could consider using a DL-based model to assist in the diagnostic process. The Laurén classification may not always be suitable for image analysis to predict the prognosis of patients. As an alternative, developing a ML-original classification could potentially yield better results. Consequently, we propose to investigate the development of such a classification system and compare its performance and similarity with conventional classification methods, including the Laurén classification. This would enable us to explore new avenues for predicting patient prognosis through image analysis, and could provide valuable insights for refining our current models.

In DL classification tasks within the field of computational histopathology, pathologist’s classification is generally treated as the ground truth when training the model [37]. Models are thus trained as best possible to imitate the pathologist’s classification. As a logical step from this, receiver operating characteristic curves and precision recall curves measuring the ability of the model to match pathologist classification are usually used as the primary form of model validation. However, in the case of certain classification tasks known to have high interobserver disagreement, one must consider carefully whether or not such methodologies are able to capture model performance accurately. Models are, as a general principle, only as good as the data they are trained, tested and validated on. There is an appreciable need for greater access and utilization of endpoints with greater objectivity such as survival, treatment response and other metrics with a lower propensity for imprecision and inaccuracies. These metrics are not without their own set of pitfalls, such a disagreement in treatment response or unclear cause of death due to a paucity of autopsies. However, these outcomes nonetheless serve to provide additional dimensions for interpreting the strengths and weaknesses of a given model. As such, when developing a model in which the ground truth is subjective and known for high levels of inter- and intra-observer disagreement, it may be worthwhile to incorporate such metrics to provide a more holistic perspective.

Our study shows that gastric adenocarcinoma subtyping on the basis of the Laurén classification can be performed using current state of the art deep learning techniques. Our DL-based classifier was able to stratify patients by survival, whereas the pathologist-based histology subtype was unable to do so. This seems to be primarily driven by the model reclassifying many resections as intestinal type which were originally classified as diffuse type. Further validation in endoscopic biopsies and detailed investigation to identify the histological survival-relevant features recognized by the DL model but not the pathologist are warranted.