MTRFU-Net: a lung nodule segmentation model based on improved U-Net architecture with spatial-frequency fusion

Accurate pulmonary nodule segmentation is a key step for the early diagnosis of lung cancer, yet existing deep learning methods still have limitations in addressing challenges such as nodule heterogeneity, blurred boundaries, and multi-scale variations. To tackle these issues, this study proposes the MTRFU-Net model, a pulmonary nodule segmentation model based on an improved U-Net architecture integrating spatial-frequency features and multi-module collaboration, which adopts a three-tier progressive module collaboration strategy to achieve dynamic spatial-frequency fusion. The encoder of the model is built on ResNet50 and incorporates a Spatial-Frequency Fusion (SFF) module, enabling the parallel extraction and dynamic fusion of dual-domain features. The bottleneck layer combines a Transformer encoder with an optimized Atrous Spatial Pyramid Pooling (ASPP) module, effectively capturing long-range dependencies and multi-scale contextual information. For the decoder, residual connections are paired with a dynamically weighted scSE attention mechanism to enhance the response capability to critical features. Extensive experiments on the LIDC-IDRI dataset demonstrate that MTRFU-Net exhibits excellent performance in terms of the Dice Similarity Coefficient (DSC), mean Intersection over Union (mIoU). This research validates the effectiveness of frequency-domain information in pulmonary nodule segmentation tasks, providing valuable references for the development of robust clinically oriented segmentation models.