Highly accurate automated tissue classification using deep learning on digital pathology images: A novel tool for resolving conflicts in diagnosis.

Abstract

3578 Background: Pathologist inspection of biopsy slides is the gold-standard for diagnosis and is crucial for effective therapy decisions. However, expert shortage is resulting in turnaround times exceeding College of American Pathologists’ (CAP) standards (Alshieban, 2015). Further, discrepancy between diagnoses can exceed 4% (Mukhopadhyay, 2018), and 2% of cases are designated as ‘carcinoma of unknown primary’ (CUP) negatively affect outcomes due to difficulty selecting therapies (Rassy, 2020). Here we sought to aid in diagnosing patients from whole-slide images (WSIs) using deep neural networks. Methods: > 6.3K high-resolution H&E-stained diagnostic WSI of formalin-fixed paraffin-embedded (FFPE) tumor block slices were selected from TCGA sources. Slide images were obtained from 30 different cancer subtypes including 368 Breast (5.6%), 324 Colon (5.12%), 287 Lung Adenocarcinoma (LUAD) (4.5%), Lung Squamous-Cell carcinoma (LUSC) (4.5%), and Stomach Adenocarcinoma (4.3%). Local regions containing tumor tissue were automatically identified by training an Inception V3 deep-learning network as previously presented. A separate Inception V3 network was trained to classify the primary tissue of 200mm2 tumor regions in 60% of the images, which was validated in the remaining 40% testing cohort. Results: The proposed deep-learning model was 92.7% accurate in identifying the primary tissue within the test set of WSIs. As expected, most misclassification occurred in highly-related tissue-types: Rectal cancers misclassified as colon (25%) and vice versa (4.8%), uveal melanomas misclassified as cutaneous melanomas (18.6%), cholangiocarcinomas as hepatocellular carcinomas (8.6%), and LUSC misclassified as LUAD (6.0%) and vice versa (3.4%). Combining related tissues, the classifier achieves 94.6% accuracy across 24 primary types. Unexpectedly, cutaneous melanomas samples were misclassified as breast (9.1%) and LUSC (5.6%), suggestive of related molecular phenotypes. Conclusions: By focusing machine-vision attention on tumor regions, the automated system approaches pathologist accuracy. Used in conjunction with molecular profiling, rates of CUP or misdiagnosis can feasibly be minimized to improve patient care. This system is currently being validated in an external set of > 4K unselected clinical cases from the NantHealth database.