ARIADNE: A Perception-Reasoning Synergy Framework for Trustworthy Coronary Angiography Analysis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Conventional segmentation methods struggle with topological consistency in coronary vessels, leading to fragmented trees.
• High pixel-level accuracy does not guarantee accurate clinical diagnosis of stenosis.

Purpose of the Study:

• To develop a novel framework (ARIADNE) for topologically consistent coronary vessel segmentation and stenosis detection.
• To improve the reliability of automated decision support systems in interventional cardiology.

Main Methods:

• ARIADNE employs a two-stage approach: preference-aligned perception using DPO (Direct Preference Optimization) and RL-based (Reinforcement Learning) diagnostic reasoning.
• The perception module fine-tunes a vision-language model with Betti number constraints for topological accuracy.
• The reasoning module uses a Markov Decision Process with an explicit rejection mechanism for reliable stenosis localization.

Main Results:

• ARIADNE achieved state-of-the-art Dice (0.8034) and centerline Dice (clDice) (0.8378).
• It demonstrated a True Positive Rate of 0.867 and reduced False Positives Per Image to 0.85 in stenosis detection.
• Outperformed generic foundation models like MedSAM3 and showed generalization on the XCAD benchmark.

Conclusions:

• Preference-based learning with structural constraints effectively mitigates topological violations in medical imaging.
• ARIADNE enhances diagnostic sensitivity and reliability in cardiology workflows, addressing alert fatigue.
• This study marks the first application of DPO for topological alignment in medical imaging.