ODySSeI: An Open-Source End-to-End Framework for Automated Detection, Segmentation, and Severity Estimation of Lesions in Invasive Coronary Angiography Images

Invasive Coronary Angiography (ICA) is the clinical gold standard for the assessment of coronary artery disease. However, its interpretation remains subjective and prone to intra- and inter-operator variability. In this work, we introduce ODySSeI: an Open-source end-to-end framework for automated Detection, Segmentation, and Severity estimation of lesions in ICA images. ODySSeI integrates deep learning-based lesion detection and lesion segmentation models trained using a novel Pyramidal Augmentation Scheme (PAS) to enhance robustness and real-time performance across diverse patient cohorts (2149 patients from Europe, North America, and Asia). Furthermore, we propose a quantitative coronary angiography-free Lesion Severity Estimation (LSE) technique that directly computes the Minimum Lumen Diameter (MLD) and diameter stenosis from the predicted lesion geometry. Extensive evaluation on both in-distribution and out-of-distribution clinical datasets demonstrates ODySSeI’s strong generalizability. Our PAS yields large performance gains in highly complex tasks as compared to relatively simpler ones, notably, a 2.5-fold increase in lesion detection performance versus a 1-3% increase in lesion segmentation performance over their respective baselines. Our LSE technique achieves high accuracy, with predicted MLD values differing by only $\pm$ 2-3 pixels from the corresponding ground truths. On average, ODySSeI processes a raw ICA image within only a few seconds on a CPU and in a fraction of a second on a GPU and is available as a plug-and-play web interface at swisscardia.epfl.ch. Overall, this work establishes ODySSeI as a comprehensive and open-source framework which supports automated, reproducible, and scalable ICA analysis for real-time clinical decision-making.

💡 Research Summary

Invasive coronary angiography (ICA) remains the clinical gold standard for assessing coronary artery disease, yet its interpretation suffers from substantial intra‑ and inter‑operator variability. To address this, the authors present ODySSeI, an open‑source, end‑to‑end framework that automatically detects lesions, segments their geometry, and estimates clinical severity metrics (minimum lumen diameter, MLD, and diameter stenosis, DS) directly from raw ICA images. The system integrates three deep‑learning components: a YOLO‑v11 based lesion detection network, a U‑Net‑style segmentation network, and a novel quantitative coronary angiography‑free lesion severity estimation (LSE) algorithm.

A central contribution is the Pyramidal Augmentation Scheme (PAS), which applies three hierarchical tiers of data augmentation during training. The static tier comprises seven domain‑specific transformations (e.g., CLAHE, median blur, motion blur, local pixel shuffling, inversion, defocus, multiplicative noise) that preserve vascular topology while expanding the dataset. The dynamic tier adds five probabilistic augmentations (random erasing, scaling, horizontal flip, translation, color jitter) to improve robustness to realistic imaging variations. Finally, a composite, scene‑based augmentation simulates occlusions and complex spatial compositions. PAS is employed for both detection and segmentation training, leading to a 2.5‑fold increase in mAP@0.50 for detection and modest (1‑3 %) gains for segmentation compared with non‑augmented baselines.

The detection model is first trained on the FAME2 dataset and further refined by incorporating the external ARCADE cohort, yielding the best performance across validation and test sets. The segmentation model receives cropped lesion patches (derived from detection) and benefits from the static and dynamic tiers of PAS. After segmentation, the LSE algorithm extracts the lesion skeleton, measures arterial radius at each skeleton point, identifies the two largest radius peaks, and defines MLD as twice the minimum radius between them. DS is computed as the complement of the MLD‑to‑maximum‑diameter ratio. This approach achieves pixel‑level MLD errors of only ±2–3 pixels relative to ground truth, demonstrating high clinical fidelity without requiring traditional QCA calibration.

Extensive experiments were conducted on three cohorts: the in‑distribution (ID) FAME2 test set and the out‑of‑distribution (OOD) Future Culprit (FC) dataset, which originates from a different patient population. While ODySSeI’s performance declines modestly on OOD data, it still outperforms baseline models, confirming the generalizability imparted by PAS. Qualitative visualizations show that the model can detect lesions missed by annotators and avoids implausible detections seen in baseline outputs.

Runtime analysis reveals that a raw ICA image is processed in a few seconds on a standard CPU and in under 0.1 s on a modern GPU, satisfying real‑time clinical requirements. The framework is made publicly available through a plug‑and‑play web interface (swisscardia.epfl.ch) and an open‑source GitHub repository (github.com/LTS4/ODySSeI), facilitating reproducibility and rapid adoption.

In summary, ODySSeI delivers a robust, accurate, and fast solution for automated ICA lesion analysis by combining hierarchical data augmentation, state‑of‑the‑art detection and segmentation networks, and a QCA‑free severity estimation method. The work addresses key barriers—subjectivity, lack of reproducibility, and computational latency—paving the way for broader clinical integration of AI‑driven coronary imaging. Future directions include multi‑view fusion, 3‑D vessel reconstruction, and extension to other cardiovascular imaging modalities such as CT and MRI.