Tailoring pretext tasks to improve self-supervised learning in histopathologic subtype classification of lung adenocarcinomas

• DOI: 10.1016/j.compbiomed.2023.107484
• Publication Date: 2023-09-16T01:56:29Z
• Authors: Ruiwen Ding; Anil Yadav; Erika Rodriguez; Ana Cristina Arau...; William Hsu
• Keywords: Digital Pathology; Subtyping

Lung adenocarcinoma (LUAD) is a morphologically heterogeneous disease with five predominant histologic subtypes. Fully supervised convolutional neural networks can improve the accuracy and reduce subjectivity of LUAD subtyping using hematoxylin eosin (H&E)-stained whole slide images (WSIs). However, developing models good prediction usually requires extensive manual data annotation, which time-consuming labor-intensive. This work proposes three self-supervised learning (SSL) pretext tasks to labeling effort. These not only leverage multi-resolution nature H&E WSIs but also explicitly consider relevance downstream task classifying Two involve predicting spatial relationship between tiles cropped from lower higher magnification WSIs. We hypothesize that these induce model learn distinguish different tissue structures presented in images, thus benefiting classification. The third involves stain stain, inducing cytoplasmic features relevant effectiveness proposed SSL their ensemble was demonstrated by comparison other state-of-the-art pretraining methods publicly available datasets. Our be extended any cancer type where architectural information important. could used expedite complement process routine pathology diagnosis tasks. code at https://github.com/rina-ding/ssl_luad_classification.