AI Sharpens Heart Disease Detection in X-Rays

Coronary artery disease is a leading cause of death worldwide, claiming millions of lives annually. Early detection is crucial for effective treatment, but standard X-ray angiograms often suffer from poor image quality, making it hard for clinicians to identify blockages accurately. Researchers have developed CASR-Net, an AI system that significantly improves the segmentation of coronary arteries in these images, offering a reliable tool to support medical diagnosis and planning.

The key finding is that CASR-Net enhances the clarity and continuity of artery structures in X-ray angiograms, achieving a Dice Score Coefficient (DSC) of 76.10% and an Intersection over Union (IoU) of 61.43%. This represents a substantial improvement over existing methods, with increases of up to 1.16% in IoU and 0.89% in DSC. The system excels at preserving narrow vessel branches and reducing false positives, which are common errors in automated analyses.

Methodologically, CASR-Net employs a three-stage pipeline. First, it preprocesses images using a multichannel approach that combines Contrast Limited Adaptive Histogram Equalization (CLAHE) and an improved version of Graham's method to enhance vessel visibility and suppress noise. Second, it segments the arteries using a deep learning model based on DenseNet121 with a Self-organized Operational Neural Network (Self-ONN) decoder, replacing traditional convolutional layers to better handle complex artery geometries. Third, it refines the results through post-processing techniques, such as contour-based removal of false positives and patch-line generation to reconnect fragmented vessel segments.

Results from a five-fold cross-validation on a combined dataset of 348 images, including both healthy and diseased arteries, show that CASR-Net outperforms state-of-the-art models. It achieved a clDice score of 79.36%, indicating strong preservation of vessel continuity. For example, in comparative tests, it reduced false negative rates and maintained high specificity, ensuring that real arteries are not missed and background noise is minimized. The system processes images quickly, with an average inference time of 0.696 seconds per image, making it practical for clinical use.

In real-world terms, this advancement matters because it can help doctors diagnose coronary artery disease more reliably, potentially leading to earlier interventions and better patient outcomes. By providing clearer images of artery blockages, CASR-Net supports treatment planning without requiring costly or invasive procedures, benefiting healthcare systems and patients alike.

However, the study has limitations. CASR-Net was not validated on external datasets, so its performance under varied clinical conditions remains uncertain. The model's computational demands may require optimization for widespread deployment, and the post-processing steps, while effective, rely partly on heuristic methods that might not adapt well to all scenarios. Future research should focus on testing the system across diverse populations and integrating temporal data from video angiograms to enhance dynamic analysis.