Reinforcement Squeeze-and-Excitation Learning for Automated Gleason Grading in Prostate Cancer Histopathology

Abstract

Proper diagnosis and treatment plans of prostate cancer can only be based on accurate Gleason grading of the prostate cancer using histopathological whole-slide images and manual assessment is likely to be subject to inter-observer variability. The present paper will suggest a new Reinforcement Squeeze-and-Excitation Learning-based Convolutional Neural Network (RIL-SE-CNN) to perform automated Gleason grading. The framework combines reinforcement learning and squeeze-and-excitation blocks to establish dynamically recalibrated channel-wise features and provides an ability to focus diagnostically relevant glandular patterns. Reinforcement feedback is an optimal method of reinforcing the attention mechanism by means of rewarding the discriminative feature representations in heterogeneous tissue regions. The proposed model is tested on two benchmark datasets, PANDA and DiagSet-A, with the help of the typical performance indicators. On the PANDA dataset, the model has an accuracy of 0.9646, precision of 0.9642, recall of 0.9657, F1-score of 0.9642 and Matthew correlation coefficient of 0.9634. The suggested method achieves better results on the DiagSet-A dataset with the following accuracy of 0.9958, precision of 0.9954, recall of 0.9957, F1-score of 0.9955, and MCC of 0.9965. These findings indicate the strength and generalization power of the suggested system in the automated Gleason grading.