Cardiac anatomical structures are highly complex, and the low contrast of computed tomography images makes accurate segmentation of the myocardium and aortic valve a challenging task. Existing U-Net-based architectures often lose high-frequency boundary information during successive downsampling operations. To address this issue, this study proposes UNetWIC, a novel dual-stream architecture that effectively preserves fine anatomical details. The encoder adopts the InceptionWT block, which consists of two branches. The spatial branch employs InceptionNeXt with large-kernel convolutions to capture global features, while the frequency branch utilizes wavelet convolution to decompose features through discrete wavelet transform and generate frequency-aware attention maps to further enhance spatial features. The decoder integrates a residual coordinate attention block to progressively refine features and accurately guide spatial reconstruction. This strategy not only effectively restores fine anatomical structures but also ensures precise boundary delineation. Five-fold cross-validation was conducted on the M-WHS-100 and MM-WHS datasets. Experimental results demonstrate that UNetWIC outperforms existing state-of-the-art models across multiple evaluation metrics, particularly in the segmentation of fine anatomical structures, thereby demonstrating its high segmentation accuracy for clinical applications.
