Deep Learning-based Unsupervised Morphological Subtyping in Histopathology Images of Gastric Cancer
- Sep. 2020
- by Kyu-Hwan Jung et. al.
Background: Gastric cancer subtyping system such as Lauren classification is widely used in clinical practice since it is an independent prognostic factor, and also provides the basis for individualized treatment. However, Lauren classification showed low reproducibility because it is based on qualitative method. Thus, a more reproducible system has been required. In this study, we explore the concordance between a cancer detection model (CDM)-based unsupervised subtyping and a current classification system.
Methods: We focused on Lauren classification since it has relatively simple major subtypes: intestinal and diffuse types. 88 resection cases of gastric cancer are collected. Two experienced gastrointestinal pathologists evaluated each case independently by Lauren’s criteria. 200 cancer region image patches per case and its features extracted by the CDM were clustered into five groups via k-means clustering. We then defined the ratio of the number of patches in each cluster group as the case-level feature. Cases are clustered into two groups via k-means clustering using the features. We named this case-level clustering result as a deep learning-based subtyping (DLS), hereafter.
Results: DLS showed a high correlation with Lauren classification: intestinal and diffuse subtype. The kappa scores between the pathologists and algorithm (0.765 and 0.796) were higher than between pathologists (0.700). The cluster centers corresponding to intestinal and diffuse types were [44.6, 21.5, 47.5, 63.7, 22.7] and [35.5, 148.2, 4.0, 9.0, 2.3], respectively. Further review of the patches contributing to the second feature (21.5 and 148.2) confirmed that they demonstrate typical morphological characteristics of diffuse type.
Conclusions: Our analysis showed that the morphological characteristics of Lauren classification were inherent in gastric cancer detection task and WSIs can be subtyped by the data-driven, and unsupervised manner. Furthermore, DLS was more reliable than human from the concordance results. This suggests that the DLS can supplement current classification systems or be a new subtyping system.
Kyu-Hwan Jung, Jeonghyuk Park, Kyungdoc Kim, Yeong Won Kim, Hyunho Park, Myeong-Cherl Kook, and Dong-Il Kim