Prediction Multiscale Cross-Level Fusion U-Net with Combined Wavelet Convolutions for Thyroid Nodule Segmentation

The precise segmentation of thyroid nodules in ultrasound images is essential for computer-aided diagnosis and treatment. Although various deep learning methods have been proposed, similar intensity distributions and variable nodule morphology often lead to blurred segmentation boundaries and missed detection of small nodules. To address this problem, we propose a multiscale cross-level fusion U-net with combined wavelet convolutions (MCFU-net) for thyroid nodule segmentation. Firstly, the network designs a multi-branch wavelet convolution (MBWC) block, which decouples texture features through wavelet domain multiresolution analysis and reorganizes cross-channel features, thereby enhancing context extraction and aggregation capabilities during the encoding stage. Secondly, a scale-selective atrous pyramid (SSAP) module based on multi-level dynamic perception is constructed to achieve saliency enhancement for nodules of varying sizes, in order to improve the detection ability for small nodules. Thirdly, to decrease the loss of fine-grained information during upsampling, a cross-level fusion module (CLFM) with hierarchical refinement mechanisms is designed, which progressively reconstructs ambiguous boundary areas through multistage upsampling. Experiments conducted on two public ultrasound datasets, TN3K and DDTI, demonstrate the effectiveness and superiority of our method, achieving Dice coefficients of 85.22% and 78.21% and IoU values of 74.25% and 64.23%, respectively.