MedFreq-Net: Medical frequency network with hybrid CNN-transformer and enhanced features for 3D medical image segmentation

3D medical image segmentation is challenging because of blurred boundaries, complex tissue structures, and scale variations, making it difficult for convolutional neural networks (CNNs) to capture both fine details and the global context. To address this challenge, we leverage both spatial and frequency-domain information to enhance the segmentation performance. We introduce MedFreq-Net, a CNN-transformer hybrid that integrates frequency-aware modules. MedFreq-Net adaptively captures information at different scales and receptive fields in the encoder and decoder to enhance feature extraction. Additionally, to address the limitation of skip connections, which typically ignore frequency-domain information, we separate frequency bandwidths and emphasize their contribution to feature representation. Furthermore, we perform frequency-domain processing on the features and integrate the information extracted by different levels of decoders through multiscale fusion. This approach effectively restores spatial details and semantic structures. We conduct experiments on four publicly available datasets and achieve promising results. Specifically, on the brain tumour segmentation dataset, MedFreq-Net achieves a Dice similarity coefficient (DSC) of 84.44%. These results indicate that MedFreq-Net significantly improves segmentation accuracy and boundary delineation for complex lesion structures. With further validation, MedFreq-Net could provide a valuable tool for clinical applications, supporting more reliable medical diagnosis and treatment planning.