Histopathology slide image classification in consideration of uncertainty in deep learning

Abstract

Convolutional Neural Networks (CNN) has been used to identify medically significant regions and objects in Whole Slide Images (WSI). Existing CNN-based studies for WSI classification showed high performance, but did not carefully consider uncertainties. Therefore, to support the accurate and efficient decision-making of pathologists, this dissertation defines and classifies uncertain situations that occur in the patch-based WSI classification process, and proposes a framework that includes methods for controlling uncertain situations. The first study focuses on how to train patch-level classifiers because patch-level predictions are used as an input for slide-level classification. The first study proposes a method to train a CNN model that classifies patches split from a WSI, considering the characteristics of histopathology images and uncertainties. The proposed method in the first study generates and uses a new sub-training dataset, which consists of mixed-patches and their new ground-truth labels, for every single mini-batch. Mixed-patches are generated using small size clean patches that have been confirmed by pathologists and their ground-truth labels are defined using a proportion-based soft labeling method. Our results obtained using a large histopathological image dataset show that the proposed method alleviates overconfidence more effectively, and performs better, than other state-of-the-art competing methods. The second study proposes a WSI classification framework consisting of solution methods for uncertain situations. To alleviate uncertain situations in the framework, we applied appropriate methods, namely noise filtering, data augmentation using mixed images, priority-based label smoothing, and a novel CNN-based slide level classification method using a feature cube. Our results show that the proposed framework control uncertain situations more effectively and performs better than other state-of-the-art competing methods on a large histopathological WSI dataset. Overall, we believe that the proposed framework will play an important role, especially in the medical domain, where uncertainty in prediction should be considered. Also, we expected that the proposed framework could be used to support accurate and efficient decision-making of pathologists.