A Generalized Deep Learning Framework for Whole-Slide Image Segmentation and Analysis

Histopathology tissue analysis is considered the gold standard in cancer\ndiagnosis and prognosis. Given the large size of these images and the increase\nin the number of potential cancer cases, an automated solution as an aid to\nhistopathologists is highly desirable. In the recent past, deep learning-based\ntechniques have provided state of the art results in a wide variety of image\nanalysis tasks, including analysis of digitized slides. However, the size of\nimages and variability in histopathology tasks makes it a challenge to develop\nan integrated framework for histopathology image analysis. We propose a deep\nlearning-based framework for histopathology tissue analysis. We demonstrate the\ngeneralizability of our framework, including training and inference, on several\nopen-source datasets, which include CAMELYON (breast cancer metastases),\nDigestPath (colon cancer), and PAIP (liver cancer) datasets. We discuss\nmultiple types of uncertainties pertaining to data and model, namely aleatoric\nand epistemic, respectively. Simultaneously, we demonstrate our model\ngeneralization across different data distribution by evaluating some samples on\nTCGA data. On CAMELYON16 test data (n=139) for the task of lesion detection,\nthe FROC score achieved was 0.86 and in the CAMELYON17 test-data (n=500) for\nthe task of pN-staging the Cohen's kappa score achieved was 0.9090 (third in\nthe open leaderboard). On DigestPath test data (n=212) for the task of tumor\nsegmentation, a Dice score of 0.782 was achieved (fourth in the challenge). On\nPAIP test data (n=40) for the task of viable tumor segmentation, a Jaccard\nIndex of 0.75 (third in the challenge) was achieved, and for viable tumor\nburden, a score of 0.633 was achieved (second in the challenge). Our entire\nframework and related documentation are freely available at GitHub and PyPi.\n