High‐Frequency Enhanced Multi‐Scale Fusion Network Using Bilateral Filtering for Cardiac Substructure Segmentation

ABSTRACT Accurate extraction of cardiac substructures from CT images plays a pivotal role in the early diagnosis and long‐term treatment of cardiovascular diseases. Although deep learning has significantly advanced cardiac segmentation, challenges persist in delineating fine substructures such as vessels. Most traditional methods neglect noise suppression and overlook the loss of high‐frequency details, leading to blurred segmentation of complex, delicate structures. As vascular substructures are distributed across different cross‐sectional layers with varying characteristics, global contextual information becomes especially important for precise segmentation. To address these issues, we propose a high‐frequency enhanced multi‐scale fusion network for cardiac substructure segmentation known as AEMF‐Net. Specifically, we introduce two modules into the AEMF‐Net architecture. First, a high‐frequency feature enhancement module is proposed, which employs bilateral filtering to enhance edge details and structural integrity in high‐frequency regions while dynamically generating pixel‐level weights to emphasise key regions and suppress redundancy. This effectively alleviates the issue of weak fine structures being overwhelmed by noise, which may lead to boundary discontinuities, as well as the loss of high‐frequency details caused by repeated pooling operations. Second, an attention‐guided multi‐scale feature fusion decoder is proposed. Using the output of the preceding encoder as a gating signal, it adaptively fuses upsampled and downsampled features, thereby enhancing the utilisation of global multi‐scale contextual information and improving segmentation accuracy for fine structures. Experimental results on a cardiac CT dataset containing 127 cases demonstrate that our method outperforms current state‐of‐the‐art approaches in segmenting ten manually annotated substructures. The proposed method achieves superior performance in terms of Dice similarity coefficient, recall, precision, and Hausdorff distance, especially in capturing complex anatomical details. The Dice coefficient for vascular substructures is improved by more than 10–30%, yielding segmentation results with enhanced clinical usability. These high‐fidelity reconstructions provide a three‐dimensional anatomical foundation for personalised cardiac modelling and surgical planning.