CCANet: Classifcation of Colorectal Tumor Histopathological Images Using a CNN with Channel Attention Mechanisms

• Main Authors: Zhang, Licheng, Cao, Fakun, Cao, Jing, Zhu, Beibei, Li, Sheng, He, Xiongxiong
• Format: Conference Proceeding
• Language: English
• Subjects: Channel Attention Mechanisms; Colorectal Tumor; Control systems; Convolutional neural networks; Dimensionality reduction; Histopathological Images Classification; Imaging; Interpretability; Learning systems; Lightweight; Pathology; Visualization
• ISSN: 2767-9861
• DOI: 10.1109/DDCLS55054.2022.9858449

Histopathological imaging analysis of colorectal polyps is the gold standard for tumor diagnosis. Nevertheless, the diagnostic process of organizing images is time-consuming and laborious. So a fast and effective computer automatic analysis method is needed. In recent years, Convolutional Neural Networks (CNN) have been frequently used for cancer classification tasks. Nevertheless, the potential global and channel relationships of the image may be overlooked and influence the representation capability of features.Also, the results are often not convincing for pathologists due to the un-interpretability. We propose a lightweight colorectal tumor classification network with only 10.02M parameters based on CNN and Channel Attention Mech-anisms(CCANet). The four classification tasks performed on the public dataset Warwick_qu reached an accuracy of 80.49%. Moreover, the visual analysis of Grad_CAM rationalizes the results and makes them more convincing to pathologists.