U-NET MEDICAL IMAGE SEGMENTATION STUDY BASED ON CBAM ATTENTION MECHANISM

In order to solve the problems of blurred lesion boundaries, missed small targets and sensitivity to background interference in 2D medical image segmentation of traditional U-Net, an improved model CB-V-UNET is proposed that fuses channel and spatial attention mechanism. Firstly, the model embeds the Convolutional Block Attention Module (CBAM) in the jump connection path of the standard U-Net to enhance the response of key areas and suppress irrelevant background noise through adaptive weighted channel features and spatial position information. Secondly, relying on the attention enhancement structure of the model, the weak boundary feature capture ability and small target recognition accuracy are optimized to make up for the structural defects of the original U-Net. Finally, model training and validation experiments were carried out on the ISIC 2018 skin lesion segmentation dataset and the DRIVE retinal vascular segmentation dataset, and the results showed that the Dice coefficient of CB-V-UNET on the ISIC 2018 dataset was 85.63%, the average intersection and union ratio (mIoU) was 85.41%, and the Dice coefficient on the DRIVE dataset was 82.15%, which was better than the original U-Net and various mainstream variants. The ablation experiment further confirmed the effectiveness of the CBAM module. This scheme provides an effective solution to improve the segmentation accuracy of small targets and weak boundary structures in medical images, and has important clinical application value.