Refined Attention Module for WSI Cancer Diagnosis

Abstract

In clinical pathology, the accurate and precise classification of cancer is one of the critical challenges due to the complex pattern of cancer cells. This study proposes an effective attention module to highlight the most important parts of cancer whole slide images (WSIs). Our attention module improves the significance of learnable features and overcomes the noisy features while training. Conventional attention modules use only feature extraction capability to learn information but we merge noisy removal capability within our attention module to leverage and randomly discard the noise during the training of the model which enhances the performance of the WSI classification task. We evaluated the performance of our module on our biopsy needle WSIs dataset, named bnWSIs. Our dataset contains a total of 24,613 labeled patches extracted from 21 WSIs. The dataset is split into two types of classification categories, with different variants of magnifications, and classes. The key is to improve the existing state-of-the-art (SoTA) performance by using the attention module, For binary classification, the achieved accuracies are improved up to 7%, whereas in multi-class classification are 6% with three magnification levels.