3D UNet-based abdominal organ automatic segmentation model using image denoising and point cloud fusion

Denoising and segmentation of medical images play an important role in early diagnosis and accurate treatment of diseases, but the existing methods still have the problems of insufficient detail preservation and limited segmentation accuracy when dealing with medical images with high noise and complex anatomical structure. In this paper, we propose an innovative framework that combines a new LKSVD denoising algorithm with image retrieval depth (RGB-D) to generate a point cloud network and improve 3D UNet network (DMSC-UNet). LKSVD significantly improves the denoising performance through sparse coding and adaptive regularization parameter prediction. By introducing dual attention module and multi-scale feature extraction module, DMSC-UNet improves the modeling ability of complex medical images and the precision of small target segmentation. The experimental results show that compared with traditional K-SVD and deep learning methods, the PSNR and SSIM indexes of LKSVD are improved by 3.03 dB and 16.7%, respectively. In the segmentation task, the average Dice score of DMSC-UNet in multiple organs reached 86.36%, which was 1.77% higher than that of 3D nnUNet. This framework effectively improves the image quality and segmentation accuracy of abdominal organs, and provides important technical support for the early diagnosis and treatment of digestive system diseases.