M2CA-Net: Multi-scale and multi-frequency channel attentional neural network for invasive coronary angiography segmentation

Accurate segmentation of invasive coronary angiography (ICA) images is crucial for diagnosing of coronary artery disease (CAD). While existing deep learning-based segmentation models have shown promising results, most operate solely in the spatial domain and overlook informative cues available in the frequency domain. To address this limitation, we design a multi-scale and multi-frequency channel attention neural network (M2CA-Net), which fuses spatial and frequency information to enhance ICA image segmentation. Specifically, we introduce a multi-frequency channel attention (MCA) block based on 2D discrete cosine transform (2D DCT) to extract global frequency representations, enhancing channel discrimination. Combined with multi-scale convolutions, this design facilitates effective fusion of spatial and frequency-domain features. We validate our model on both public and clinical datasets, where M2CA-Net achieves superior segmentation performance and outperforms several state-of-the-art architectures.