DRLSU-Net: Level set with U-Net for medical image segmentation

Convolutional neural networks (CNN) have been extensively utilized for image segmentation tasks, with the U-Net architecture emerging as a classical model in medical imaging due to its simple structure and high scalability. However, for complex medical images, particularly those with blurred lesion boundaries, the U-Net model often loses significant edge information during feature extraction. Each layer in the encoder section is convolved with a simple stack of equal design, which is clearly not able to obtain sufficient feature information from an otherwise low-quality image. In order to solve this problem, the DRLSU-Net model is proposed which combines an enhanced U-Net architecture with distance regularized level set evolution (DRLSE). The DRLSU-Net model takes the results of U-Net pre-segmentation as an intermediate medium, combines the U-Net model with the level set method. Indirectly represent the target contour through the zero level set to obtain a more intuitive target edge location. Specifically, the Parallel Dilated Convolutional Sequence (PDCS) is introduced in the U-Net encoder to minimize information loss during down-sampling, and preserve more edge details. Secondly, the Mixed Attention Mechanism (MAM) is introduced into the decoder, aiding the network in recovering important information during image reconstruction, thus generating a more accurate output sequence. Finally, the pre-segmentation label mapping is converted into a level set function representation, which serves as a priori information for the level set method. A new energy functional is constructed to guide the evolution of the level set curves, helping to obtain a clear contour boundary. The performance of the DRLSU-Net model is evaluated on the ISIC2017, ISIC2018, CVC-ClinicDB, and Lung datasets. Extensive experiment results show better performance than other state-of-the-art (SOTA) methods in terms of mIoU and F1-socre, and the results indicate that the DRLSU-Net model performs competitively in medical image segmentation tasks.