Computerized diagnosis of prostate cancer based on whole slide histology images and deep learning methods

Abstract

This paper presents the results of an experimental study and the development of tools for automatic analysis and recognition of histological images in order to obtain quantitative estimates of the presence and degree of aggressiveness of prostate cancer in the commonly used Gleason and ISUP scales. The input data consisted of 10 616 whole-slide histological images with the size of the largest side up to 100 000 pixels and

22 089 of their image tiles of 256×256 pixels in size. Two solutions were chosen as the final ones. The first solution is based on sequential analysis of image fragments and includes feature extraction using the ResNet50 network and the subsequent generalization of particular recognition results using a small convolutional network. The second solution is based on the simultaneous analysis of the selected informative sections, presented in the form of an intermediate pseudo-image, and its subsequent recognition using an ensemble of four variants of convolutional networks with the EfficientNetB0 architecture. Being independently tested on an unknown image dataset that was not available for authors, these approaches achieved the prediction accuracy of 0,9277 according to the ISUP scale.