Polyp image segmentation based on parallel dilated convolution and dual attention mechanisms.

Colorectal cancer is usually caused by malignant transformation of early colon polyps. Early polyps are benign, but if left untreated, they can progress to cancer. Early diagnosis of polyps can significantly improve the survival rate of patients. To overcome the challenges of small polyp localization and the blurring of segmentation boundaries caused by indistinct polyp edges in existing research on colon cancer medical image diagnosis, we propose MSFNet (Muilti-Scale Feature Unet), a polyp image segmentation model using multi-scale contextual feature extraction and re-aggregation channel attention modules. MSFNet extracts detailed information separately from spatial and channel features, which can efficiently explore image details and help locate small polyp positions. Simultaneously using channel by channel convolution and dilated convolution while maintaining a low parameter count, MSFNet can achieving strong performance on small datasets. In addition, we propose MBDCK (Multi Scale Big Dilated Conv kernel) module, which consists of dilated large convolutions with different kernel sizes in parallel, to enhance the multi-scale context information and solve the limited receptive field problem. We design skip connections to fuse feature map information at multiple scales, which enhance the model's ability to obtain global features and suppresses the extraction of irrelevant features. Theoretical analysis and extensive experimental results demonstrate that the method proposed in this paper, compared to existing advanced medical image segmentation methods, has fewer parameters while achieving excellent segmentation accuracy and generalization ability. On CVC-ClinicDB, MSFNet attains a Dice of 0.892 and an mIoU of 0.926, out-performing U-Net by about 10% in mIoU and surpassing recent advanced models such as PraNet and SwinUNet.