WCE-Bleed-Net: A Convolutional Neural Network with Adaptive Attention Module for Bleeding Frame and Region Detection in Wireless Capsule Endoscopy Video

Wireless capsule endoscopy (WCE) offers a non-invasive and comfortable approach of inspecting a patient's whole gastrointestinal (GI) tract by direct observation. The identification of bleeding frames in WCE video and segmenting bleeding regions with in the frame with irregular shaped bleeding regions can be challenging due to the varied intensity, intricate distribution of the bleeding areas with differing morphologies. Particularly, when the bleeding region is small and colour intensity distribution is same as surrounding regions. n this study, we introduce a deep convolutional neural network (CNN) model integrated with an adaptive attention mechanism. This model incorporates spatial and channel attention modules within its convolutional layers, enabling it to extract features across varying receptive fields. The attention mechanism allows the network to adaptively focus on robust representations in both spatial and channel dimensions, making it more effective at handling complex bleeding patterns. The proposed framework includes two specialized CNN models: one designed to detect bleeding frames in W CE videos and another to accurately locate bleeding regions within those frames. The models were trained from scratch using the Auto-WCEBleedGen dataset, and their performance was rigorously compared with existing methods. Experimental results highlight the advantages of the proposed WCE-Bleed-Net over recent classification approaches and U-Net-based segmentation techniques, demonstrating its effectiveness in addressing the challenges of bleeding detection and segmentation in WCE data.