Deep learning models based on post-procedural angiography for predicting the future risk of in-stent restenosis

In-stent restenosis (ISR) remains a significant complication after percutaneous coronary intervention. Identifying patients at high risk for ISR from visually patent post-procedural angiograms is challenging. We developed and validated deep learning models using routine post-procedural digital subtraction angiography (DSA) to predict future ISR risk, emphasizing the role of mask-guided spatial attention. We retrospectively enrolled 237 patients with 1-year angiographic follow-up across three centers. Immediate post-procedural DSA frames were used as inputs, with 1-year outcomes as ground truth. Deep learning architectures were trained using full-image and mask-guided strategies. On the independent external test set, the mask-guided DenseNet-121 achieved the highest predictive performance (AUC: 0.885, AUPRC: 0.912). Crucially, models trained on full images exhibited severe shortcut learning, erroneously focusing on task-irrelevant background noise. The mask-guided strategy successfully corrected this, demonstrating excellent calibration and robust clinical net benefit. Grad-CAM visualization confirmed precise attention on stented vessel regions associated with future restenosis. In conclusion, mask-guided deep learning analysis of post-procedural DSA provides an accurate, interpretable, and automated prognostic tool. By overcoming shortcut learning, this approach effectively stratifies future ISR risk, holding potential to guide personalized post-procedural surveillance and optimize secondary prevention.Trial registration: This study was registered with the Chinese Clinical Trial Registry (ChiCTR) under the registration number ChiCTR2500115474 on December 26, 2025.