Unsupervised, Self-supervised, and Supervised Learning for Histopathological Pattern Analysis in Prostate Cancer Biopsy

Abstract

Prostate cancers are often non-aggressive, which makes it difficult to determine a treatment such as surgical prostate removal. The survival rate can be significantly enhanced with early detection of cancer so that appropriate intervention can be administered. This paper presents a state-of-the-art system that conducts a multi-class classification, from unsupervised to supervised learning techniques to predict the tissue components, namely stroma benign, and cancer in whole slide image (WSI) of prostate cancer. First, the unsupervised classifier learns from an unlabeled dataset to extract meaningful information from tissue images. For that, we used a modified K-means algorithm without any supervision to generate the labels using scale-invariant feature transform (SIFT), Histogram of oriented gradients (HOG), Gray-level co-occurrence matrix (GLCM), and Edge-based (i.e., Sobel, Roberts, Scharr, and Prewitt) features. Further, our proposed model, Bi-directional ConvLSTM Convolutional Neural Network (BCACNN) is used for self-supervised and supervised learning on unlabeled and unsupervised labeled data, respectively, to differentiate the regions in WSI of prostate cancer. The proposed model achieved the highest accuracy, precision, recall, f1-score, and area under the curve (AUC) of 0.8655, 0.8597, 0.8524, 0.8560, and 0.9748, respectively, on the internal test dataset, and 0.8915, 0.8880, 0.8904, 0.8891, and 0.9764, respectively, on the external test dataset.