EFAG-CNN: Effectively fused attention guided convolutional neural network for WCE image classification

Wireless capsule endoscopy (WCE) has been widely used in the detection of digestive tract diseases because of its painlessness and convenience. Accurate classification of WCE abnormal images is very crucial for the diagnosis and treatment of early gastrointestinal tumors, while it remains challenging due to the ambiguous boundary between lesions and normal tissues. In order to overcome the above limitations, a three-branch effectively fused attention guided convolutional neural network (EFAG-CNN) which imitates the practical diagnosis process is proposed. Specifically, global features and local images with suppressed background noise are generated by branch1 and local features are extracted by branch2 based on the local images. What's more, an effective attention feature fusion (EAFF) module is devised and inserted into branch3 to make the final prediction, which helps adaptively capture more discriminative features for classification. EAFF can integrate the representative features from branch1 and branch2 better than other methods. Furthermore, we propose a joint loss function to enhance the classification performance of branch2. Extensive experimental results demonstrate that the overall classification accuracy of the proposed method on the public Kvasir dataset reaches 96.50%, which is superior to the state-of-the-art deep learning methods.