Medical image segmentation model based on multi-scale fusion and feature reconstruction convolution

Medical image segmentation plays a vital role in assisting clinical diagnosis and treatment. However, thyroid ultrasound images often suffer from low contrast, severe noise interference, and blurred boundaries, making accurate and robust segmentation a challenging task for conventional methods. To address these limitations, we propose a novel segmentation model, CS-Net, which integrates multi-scale feature fusion with feature reconstruction convolution. The model introduces a Spatial and Channel Reconstruction Convolution module (SCConv) to suppress redundant features and enhance the extraction of fine-grained details in low-quality regions. Additionally, a Channel-wise Cross Fusion Transformer (CCT) module is incorporated to align and integrate multi-level semantic features, effectively bridging the semantic gap between encoder and decoder stages. Furthermore, a Pixel-attention-guided fusion and Upsample module (PagU) is designed to dynamically reweight fused features, improving boundary delineation and localization accuracy. Extensive experiments on two public thyroid ultrasound datasets, DDTI and TN3K, demonstrate that CS-Net outperforms representative models including UNet, TransUNet, and SwinUNet in terms of Dice coefficient, mIoU, precision, and recall. Specifically, CS-Net achieves Dice scores of 87.36 % and 86.47 % on the DDTI and TN3K datasets, showcasing excellent robustness. Ablation studies further confirm the effectiveness of each module. These results suggest that CS-Net offers a promising and reliable solution for automatic segmentation of low-quality medical images.