Training based segmentation for tissue extraction in whole slide image

Abstract

Reducing the time and storage memory required for scanning whole slide images (WSIs) is crucial. In this thesis work we tested and assessed the performance of two popular neural network architectures, namely DeepLabV3+ and Unet. In addition to that, a desktop application used to annotate histopathology images was developed, such application ultimately provided the data needed in order to train the neural networks. Both DeepLabV3+ and Unet accurately separated the regions of interest out of the WSIs, however DeepLabV3+ outperformed Unet, striking a pixel wise accuracy of 96.3%, while Unet scored 94.7% in the same metric. Morover DeepLabV3+ also outscored Unet in the IoU metric with values of 0:446 and 0:398 respectively. We showed the effectiveness of using deep neural networks for the case of semantic segmentation in histopathology images, more specifically for extracting tissue areas from WSIs, and how this can be used to improve the performance of WSI scanners.