<p>Accurately and automatically segmenting the hippocampus into multiple subfields on magnetic resonance imaging images is crucial for the diagnosis and intervention of various neurological diseases. However, small sizes and complex structures of the hippocampal subfields, along with their unclear boundaries and significant volume imbalances, pose substantial challenges for automatic segmentation. To address these challenges, we propose a novel Multi-level Wavelet Fusion Network (MWFNet) to enhance the segmentation accuracy of hippocampal subfields. MWFNet incorporates multi-level wavelet transforms during the encoding process, effectively supplementing spatial domain information with wavelet domain features to improve the perception of detailed semantic information. Additionally, it recovers information lost due to downsampling, particularly benefiting smaller hippocampal subfields. Additionally, we developed a Multi-scale Attention Residual Block (MARB) that leverages convolutional kernels of different sizes to facilitate multi-scale feature extraction. MARB integrates channel and spatial attention to adaptively extract the most effective image features. Combining MARB, we also introduced a new deep supervision scheme to enhance MWFNet’s attention and supervision on effective deep features. Extensive experiments conducted on two public hippocampal subfield datasets show that our approach surpasses other state-of-the-art methods. Specifically, MWFNet achieved an average Dice Similarity Coefficient (DSC) score of 75.27% and an average 95th percentile Hausdorff Distance (HD95) score of 0.96 mm on the PHS dataset; on the UMC dataset, it achieved an average DSC score of 77.66% and an average HD95 score of 1.08 mm. Compared to existing hippocampal subfield segmentation methods, MWFNet represents a significant attempt to incorporate multi-level wavelet transforms into this task and has illustrated superior performance.</p>