CASR-Net: An Image Processing-focused Deep Learning-based Coronary Artery Segmentation and Refinement Network for X-ray Coronary Angiogram

Early detection of coronary artery disease (CAD) is critical for reducing mortality and improving patient treatment planning. While angiographic image analysis from X-rays is a common and cost-effective method for identifying cardiac abnormalities, including stenotic coronary arteries, poor image quality can significantly impede clinical diagnosis. We present the Coronary Artery Segmentation and Refinement Network (CASR-Net), a three-stage pipeline comprising image preprocessing, segmentation, and refinement. A novel multichannel preprocessing strategy combining CLAHE and an improved Ben Graham method provides incremental gains, increasing Dice Score Coefficient (DSC) by 0.31-0.89% and Intersection over Union (IoU) by 0.40-1.16% compared with using the techniques individually. The core innovation is a segmentation network built on a UNet with a DenseNet121 encoder and a Self-organized Operational Neural Network (Self-ONN) based decoder, which preserves the continuity of narrow and stenotic vessel branches. A final contour refinement module further suppresses false positives. Evaluated with 5-fold cross-validation on a combination of two public datasets that contain both healthy and stenotic arteries, CASR-Net outperformed several state-of-the-art models, achieving an IoU of 61.43%, a DSC of 76.10%, and clDice of 79.36%. These results highlight a robust approach to automated coronary artery segmentation, offering a valuable tool to support clinicians in diagnosis and treatment planning.

💡 Research Summary

The paper introduces CASR‑Net, a three‑stage deep learning framework designed to automatically segment coronary arteries from X‑ray coronary angiograms, a task that is challenged by low contrast, noise, and the presence of narrow, stenotic vessel branches. The first stage employs a novel multichannel preprocessing pipeline that fuses Contrast Limited Adaptive Histogram Equalization (CLAHE) with an improved Ben Graham filter. CLAHE enhances local contrast while the Ben Graham filter simultaneously reduces noise and sharpens edges. By concatenating the two processed images along the channel dimension, the authors achieve incremental improvements over using either technique alone, reporting a Dice Score increase of 0.31–0.89 % and an IoU gain of 0.40–1.16 %.

The core segmentation network builds upon the classic UNet architecture but replaces the standard encoder with a pre‑trained DenseNet‑121, leveraging its dense connectivity to capture rich hierarchical features without excessive parameter growth. The decoder is replaced by a Self‑organized Operational Neural Network (Self‑ONN), which learns a set of adaptable operators rather than fixed convolution kernels. This flexibility enables the model to preserve continuity in very thin or highly curved vessel segments, a common failure mode for conventional convolutional decoders. Skip connections from the encoder are retained, ensuring high‑resolution spatial information is available throughout the up‑sampling path.

A final refinement module addresses residual false positives. It applies morphological opening to eliminate small isolated blobs, followed by connected component analysis that discards components below a predefined area threshold. To explicitly encourage topologically correct vessel structures, the authors incorporate clDice—a loss term that measures overlap of connected components—into the training objective. This post‑processing step yields an additional 1–2 % boost in IoU and Dice metrics.

For evaluation, the authors combined two publicly available coronary angiography datasets, encompassing both healthy and diseased cases, and performed 5‑fold cross‑validation. They benchmarked CASR‑Net against several state‑of‑the‑art segmentation models, including vanilla UNet, Attention‑UNet, DeepLabV3+, SegFormer, and recent transformer‑based approaches. CASR‑Net achieved an average Intersection over Union of 61.43 %, a Dice Score of 76.10 %, and a clDice of 79.36 %, surpassing all competitors. Notably, the improvement was most pronounced in stenotic regions where preserving narrow branch continuity is critical.

The authors highlight four main contributions: (1) a dual‑technique multichannel preprocessing that jointly enhances contrast and suppresses noise; (2) a DenseNet‑121 encoder paired with a Self‑ONN decoder that maintains thin‑vessel continuity while remaining parameter‑efficient; (3) a contour refinement stage that reduces false positives and enforces topological consistency via clDice; and (4) extensive validation on a merged dataset that demonstrates superior performance over existing methods.

Limitations include the added computational overhead of the preprocessing step and the sensitivity of Self‑ONN training to hyper‑parameter choices. Future work will focus on streamlining the preprocessing pipeline, exploring more generalized operational layers, and extending the approach to other vascular imaging modalities such as CT‑angiography and MR‑angiography. In summary, CASR‑Net offers a robust, clinically relevant solution for automated coronary artery segmentation, potentially aiding cardiologists in diagnosis, risk stratification, and treatment planning.