Automatic Gleason Classification of Prostate Cancer - Classification of Small Regions

Abstract: Purpose: To classify the severity of a case of prostate cancer, physicians use the 10-grade Gleason score. The purpose of this Master’s thesis is to study how small dimensions of image crops affect the Gleason 5-classification capability of a machine learning system. In this thesis, two aspects of dimensionality have been taken into account when creating image crops, the image crop size and the degree of magnification. Methodology: 70 x 70 and 128 x 128 pixel images, both with a 40X magnification, were cropped from larger tissue images annotated at Skåne University Hospital (SUS), creating one data set for each image crop size. The networks trained on these data sets were as follows: a CNN-architecture, a CNN-architecture with an Inception-v4-module at the end, a ResNet-architecture, and a CNN-architecture with an Inception-ResNet-v1-module at the end. Results: The ResNet-architectures performed the best on the created data sets, achieving mean 5-fold cross-validation accuracies of 91.9% and 96.5 % for the 70 x 70 and 128 x 128 pixel images respectively. However, these architectures experienced temporary drops in accuracy. Furthermore, the modified CNN-networks could not be determined to definitely outperform the base CNN-networks. Conclusion: The results indicated that image crops of sizes larger than 70 x 70 when using a magnification of 40X were preferable for PCa-classification purposes. However, the classification effects of using different architecture designs were inconclusive.