International Conference on Computer and Knowledge Engineering

Home / 12th International Conference on Computer and Knowledge Engineering

Weakly Supervised Convolutional Neural Network for Automatic Gleason Grading of Prostate Cancer

Authors :

Maryam Kamareh¹ Mohammad Sadegh Helfroush² Kamran Kazemi³

1- Department of Electrical and Electronics Engineering, Shiraz University of Technology, Shiraz, Iran 2- Department of Electrical and Electronics Engineering, Shiraz University of Technology, Shiraz, Iran 3- Department of Electrical and Electronics Engineering, Shiraz University of Technology, Shiraz, Iran

Keywords :

Histopathology،Prostate cancer،Gleason Grading،Deep Learning،Weakly Supervised،Convolutional Neural Network

Abstract :

Digital histopathology is based on the analysis of digitized biopsy slides. Prostate cancer is the second most common cancer among men. The grading of prostate cancer is based on the Gleason grading system. The process of analyzing histopathology images and manually determining the Gleason grades by expert pathologists is a time-consuming and expensive process. Therefore, development of a system based on machine learning can provide an accurate method for grading prostate cancer. But these systems are not easy to develop because they require significant amounts of pixel level annotated data. Since the pathologists' clinical reports often contain only slide-level labels, this type of data is rarely available. Therefore, the development of methods that can learn using only slide-level labels and do not require manual pixel level annotation would be a significant advance in this field. In this paper, we propose a weakly-supervised convolutional neural network for Gleason grading of prostate cancer tissue microarrays with Hematoxylin and Eosin staining, without using pixel level annotations. To choose a suitable classifier, we explored different pre-trained models as a feature extractor, namely ResNet-50, VGG-19 and MobileNet. We used class-wise data augmentation method to face the imbalance problem in our dataset. The best network architecture was ResNet-50 as backbone. In the test cohort, ResNet-50 as backbone with class-wise data augmentation achieved an accuracy of 80% for the Gleason grades.