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

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Interventional Cardiology

Background:

• In-stent restenosis (ISR) is a major complication following percutaneous coronary intervention.
• Predicting ISR risk from visually normal angiograms post-procedure is difficult.
• Routine digital subtraction angiography (DSA) data holds potential for risk stratification.

Purpose of the Study:

• To develop and validate deep learning models for predicting future ISR risk using post-procedural DSA.
• To investigate the role of mask-guided spatial attention in improving model accuracy and interpretability.
• To address shortcut learning issues in AI models applied to medical imaging.

Main Methods:

• Retrospective cohort of 237 patients with 1-year angiographic follow-up.
• Deep learning models (DenseNet-121) trained on immediate post-procedural DSA frames.
• Comparison of full-image versus mask-guided attention strategies.
• External validation and Grad-CAM visualization for interpretability.

Main Results:

• The mask-guided DenseNet-121 model achieved high predictive performance (AUC: 0.885, AUPRC: 0.912) on an independent test set.
• Full-image models showed significant shortcut learning, focusing on irrelevant background noise.
• Mask-guided strategy corrected shortcut learning, demonstrating good calibration and clinical utility.
• Attention maps highlighted stented vessel regions relevant to future restenosis.

Conclusions:

• Mask-guided deep learning analysis of post-procedural DSA is an accurate, interpretable, and automated prognostic tool for ISR.
• This method effectively stratifies ISR risk by overcoming shortcut learning.
• The approach can guide personalized surveillance and optimize secondary prevention strategies post-intervention.