FE-Unet: A lightweight brain tumor segmentation network combining frequency domain processing and edge preservation

Brain gliomas are the most common malignant tumors, which threaten the patients' life, so the precise segmentation of the tumor lesion is a key step in the use of computer-aided diagnosis. However, this work faces challenges due to the diversity of brain tumors and the complexity of multi-modality images. To this end, this article proposes FE-Unet, a lightweight brain tumor segmentation network that combines frequency domain processing with edge preservation. FE-Unet effectively improves U-Net regarding the encoder and extended path, and we also propose an edge-enhanced hybrid (EEH) loss function that can strengthen the components above and improve the segmentation effect for irregular edges. A feature reconstruction wavelet (FRW) encoder is designed instead of vanilla convolution to eliminate feature redundancy among multi-modality images and edge preservation. Moreover, an orthogonal attention cross-scale guided (OACG) module based on frequency domain processing is proposed to narrow the semantic gap between the two ends of the expansion path and complete the cross-scale guided up-sampling step. Compared with other advanced baseline methods, FE-Unet performer best in Dice similarity coefficient, Hausdorff distance, etc., with only 6 % of the number of parameters and 4 % of the number of multiply-accumulate operations (MACs) of the U-Net.