A Multiscale Intracerebral Hemorrhage Segmentation Framework Based on Wavelet Convolution and Feature Compensation Mechanism

Abstract

Artificial intelligence–assisted diagnostic technologies are increasingly applied in the medical field, particularly for the diagnosis of intracerebral hemorrhage (ICH), a cerebrovascular disease with high mortality and disability rates. Although CT imaging is the primary modality for ICH diagnosis, accurate and rapid lesion segmentation remains challenging because hemorrhagic lesions are small, spatially variable, and heterogeneous in appearance, and conventional models are limited in detecting small lesions due to their small receptive fields and feature loss. To address these challenges, we propose WRes-UNet (Wavelet Convolution Residual-UNet), an enhanced U-Net–based segmentation framework that integrates Wavelet Transform Convolution (WTConv) and an Adaptive Feature Shortcut (AFS) mechanism. WTConv decomposes feature maps into multiple frequency subbands, enabling multiscale contextual representation while preserving critical low-frequency information. The AFS module adaptively enhances discriminative channel features, effectively compensating for feature loss and improving the localization of small hemorrhagic regions. Extensive experiments on an ICH CT dataset show that WRes-UNet consistently outperforms state-of-the-art segmentation models in Dice, IoU, and F1 scores, while achieving lower model complexity. In particular, the proposed framework demonstrates clear advantages in segmenting small and irregular hemorrhagic lesions. These resultsindicate that WRes-UNet provides an effective and robust solution for precise ICH segmentation, showing promising clinical value for early diagnosis. The code for WRes-UNet is available at https://github.com/GZWANGWENYU/WRes-UNet.