Improving Explainability of the Attention Branch Network with CAM Fostering Techniques in the Context of Histological Images

Abstract

Convolutional neural networks have presented significant results in histological image classification. Despite their high accuracy, their limited interpretability hinders widespread adoption. Therefore, this work proposes an improvement to the attention branch network (ABN) in order to improve its explanatory power through the gradient-weighted class activation map technique. The proposed model creates attention maps and applies the CAM fostering strategy to them, making the network focus on the most important areas of the image. Two experiments were performed to compare the proposed model with the ABN approach, considering five datasets of histological images. The evaluation process was defined via quantitative metrics such as coherency, complexity, confidence drop, and the harmonic average of those metrics (ADCC). Among the results, the proposed model through the ResNet-50 was able to provide an improvement of 4.16% in the average ADCC metric and 3.88% in the coherence metric when compared to the respective ABN model. Considering the DesneNet-201 network as the explored backbone, the proposed model achieved an improvement of 14.87% in the average ADCC metric and 9.77% in the coherence metric compared to the corresponding ABN model. The contributions of this work are important to make the results via computer-aided diagnosis more comprehensible for clinical practice.