Deep learning-based polyp image segmentation is helpful for assessing precancerous lesions. A colorectal polyp image segmentation model (RDG-Net) based on a dual-stage decoder is proposed to addresses the issues of poor segmentation performance due to unclear boundaries in colorectal polyp images and insufficient generalization ability for new samples. The model uses Res2Net-50 as the encoder to enhance the precision of image segmentation. The decoder has two stages:the first stage utilizes a 4-layer multi-scale feature aggregation module to integrate features extracted by the encoder at different stages,while the second stage enhances the image features output by the first stage of the decoder through a 3-layer parallel convolution fusion module and decodes them to a higher resolution as the model's final output. The model is trained using the CVC-ClinicDB and Kvasir-SEG training datasets,and tested using the CVC-ClinicDB and Kvasir-SEG test datasets,as well as the CVC300 and ETIS-LaribPolypDB datasets that are not involved in the training. The test results show that the proposed method has an average accuracy,precision,recall rate,Dice similarity coefficient,intersection over union and F2 score of 98.41％,94.25％,92.62％,93.42％,87.69％and 92.93％on the CVC-ClinicDB and Kvasir-SEG datasets,while 99.05％,87.79％,89.13％,88.39％,79.33％ and 88.82％ on the CVC300 and ETIS-LaribPolypDB datasets,respectively,demonstrating that RDG-Net model performs well in colorectal polyp region segmentation and has a high generalization performance on new datasets.

Deep learning-based polyp image segmentation is helpful for assessing precancerous lesions. A colorectal polyp image segmentation model (RDG-Net) based on a dual-stage decoder is proposed to addresses the issues of poor segmentation performance due to unclear boundaries in colorectal polyp images and insufficient generalization ability for new samples. The model uses Res2Net-50 as the encoder to enhance the precision of image segmentation. The decoder has two stages:the first stage utilizes a 4-layer multi-scale feature aggregation module to integrate features extracted by the encoder at different stages,while the second stage enhances the image features output by the first stage of the decoder through a 3-layer parallel convolution fusion module and decodes them to a higher resolution as the model's final output. The model is trained using the CVC-ClinicDB and Kvasir-SEG training datasets,and tested using the CVC-ClinicDB and Kvasir-SEG test datasets,as well as the CVC300 and ETIS-LaribPolypDB datasets that are not involved in the training. The test results show that the proposed method has an average accuracy,precision,recall rate,Dice similarity coefficient,intersection over union and F2 score of 98.41％,94.25％,92.62％,93.42％,87.69％and 92.93％on the CVC-ClinicDB and Kvasir-SEG datasets,while 99.05％,87.79％,89.13％,88.39％,79.33％ and 88.82％ on the CVC300 and ETIS-LaribPolypDB datasets,respectively,demonstrating that RDG-Net model performs well in colorectal polyp region segmentation and has a high generalization performance on new datasets.