HFSUNet: A UNet architecture optimized for precise segmentation of pneumonia lesion boundaries

Pulmonary lesion segmentation, a key component of computer-aided diagnosis, has significant clinical value in accurately localizing infected regions, quantifying disease progression, and guiding individualized treatment plans. To address the limitations of traditional methods-reliance on handcrafted features, vulnerability to medical image noise, and susceptibility to physician subjectivity-this study proposes HFSUNet, an improved UNet-based segmentation model. The model introduces the multi-scale attention module (MSAM) module to capture cross-scale salient lesion features (alleviating boundary blurring and detail loss), Haar wavelet downsampling (HWD) to reduce downsampling information loss, and marginal weight loss (MWL) to optimize edge segmentation affected by blurred lesion boundaries. Experimental results show HFSUNet achieves mIoU scores of 99.75% (COVID-19), 86.91% (MosMedData), and 92.77%, 92.28%, and 94.16% on DDTI, BUSI, and TN3K, outperforming UNet-based baselines. Notably, while maintaining high segmentation accuracy, it only has 5.638 million parameters and 11.582 GFLOPs, significantly reducing complexity and computational cost. This advantage enables clinical application in resource-limited settings, demonstrating great potential in lesion identification, diagnostic assistance, and prognosis support.