Mitigating error propagation in whole slide images analysis using deep learning

Abstract

With the advancements of deep learning, specifically convolutional neural networks (CNNs), researchers utilize CNNs for classifying the whole slide images. To make use of CNNs, researchers follow a two-stage scheme. First, they divide whole slide images into patches to train a CNN called patch level model, and in the second stage, they aggregate CNN's output for all patches to assign the final label called slide level model. Different forms of CNNs managed to deliver promising results highlighting the potential of deep learning. However, most of the work has been done to improve the CNNs performance using different architectures and ignored the error propagation in whole slide image analysis frameworks. This research highlights how errors can propagate between different stages of the whole slide image analysis framework and proposes methods to mitigate such errors. We have identified data-driven and model-driven errors in stage 1 and stage 2, respectively. To reduce the intensity of errors, we first proposed a loss-based method called LossDiff to address data-driven errors. Secondly, an unsupervised method WSI-GAD based on graph autoencoders is incorporated to mitigate model-driven errors. Several evaluation methods were applied, highlighting the presence of error and showing that the proposed methods significantly improve the classification performance. The dissertation provides valuable contributions to computer-aided whole slide image analysis literature. This study's findings will direct the research community to consider the uncertainty injected by the training data and understand how the model-driven errors can affect the final predictions of the whole slide image. This will help researchers to develop not only accurate models but also deliver generalized performance.