Deep learning for x-ray coronary angiography: a systematic review of methods, validation, and clinical readiness

Background Coronary artery disease (CAD) is the leading cause of death globally. While X-ray coronary angiography (XCA) guides diagnosis and percutaneous coronary intervention, interpretation remains expert-dependent.1 Ongoing digitalization introduces opportunities to embed data-driven and AI-supported intelligence into the catheterization laboratories workflow. Purpose This systematic review evaluates deep learning (DL) applications to XCA, focusing on methods, validation strategies, and indicators of translational feasibility. Methods We conducted a PRISMA-guided systematic review, searching PubMed, Web of Science, and IEEE Xplore (2014–2025) for peer-reviewed studies developing deep learning (DL) models for X-ray coronary angiography (XCA). Database searches were supplemented by forward/backward citation chasing via Lens.org. Two independent reviewers screened titles and abstracts using Rayyan, resolving conflicts with a third reviewer; full-text assessment followed identical procedures. Studies were quality-rated using CLAIM checklist.2 Translational feasibility was evaluated through external validation evidence, reference standard agreement, real-time performance metrics, generalizability across populations/equipment, and CLAIM reporting transparency. Results We included 118 studies (2016-2025), observing a surge in research output after 2020. Most studies (86%) focused on a single task, mainly vessel segmentation and CAD detection (Figure 1). While 91% specified the target vessel, population demographics were reported in 14% and patient presentation in 23%. Methodological robustness was limited: 25% of studies included external validation, 20% used multi-centre data, and none employed a prospective design. Convolutional Neural Networks (CNNs) were the most frequently utilized models. Nine studies employed self-/semi-supervised learning to address limited annotations. Transfer learning saw a steady increase (from 2% in 2019 to 11% in 2025), yet no foundational models were used. Annotator expertise was described in 77%, 29% used ≥2 experts. Reporting quality (CLAIM 2024; median = 48.8%) was moderate, with sparse reporting on ground truth (5% reported the inter/intra-rater variability) and missing data (2%). Translational feasibility indicators revealed systematic limitations across all domains (Figure 2B). Real-time feasibility was demonstrated in 19% and inference time reported in 28% of the papers. Reproducibility remained low (availability of code 28%, data 21%, models <1%). Conclusion DL for XCA is advancing rapidly, reflected in growing research output and increasing methodological sophistication, yet a clear path from technical innovation to clinical translation remains to be established. With the adoption of standardized benchmarking, open science practices, and clinically grounded frameworks such as CLAIM and FUTURE-AI, the field is well-positioned to bridge this gap.