Development of double U-Net model using depthwise separable convolution and convolutional block attention module (CBAM) for polyp medical image segmentation

Abstract

One of the most common types of cancer is colorectal cancer, which is commonly caused by colon polyps. Early detection via colonoscopy plays a crucial role in preventing the advancement of cancer, even though initial symptoms may not always be noticeable. In this context, computer-assisted methods for image segmentation in colonoscopy significantly aid in polyp detection, reduce the workload on healthcare professionals, and enhance diagnostic precision. A widely used approach in medical image segmentation is the U-Net model, known for its ability to merge information across multiple resolutions through its encoder-decoder structure. Nonetheless, several U-Net-derived architectures struggle to effectively capture deeper features and maintain computational efficiency. This research introduces a modified Double U-Net architecture that integrates MobileNetV2 as its encoder to alleviate computational complexity, while incorporating Depthwise Separable Convolutions (DWSC) and Convolutional Block Attention Modules (CBAM) in the skip connections to enhance both accuracy and efficiency. The proposed model demonstrated notable performance, achieving an F1-score of 95.28% and IoU of 90.84% on the CVC-Clinic DB dataset, and an F1-score of 89.08% with an IoU of 80.69% on the Kvasir-SEG dataset, with a parameter count of 1.96 million. This approach not only increases segmentation precision but also maintains computational efficiency, making it a viable option for detecting colorectal polyps.