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

ARIADNE is a novel two-stage framework that enhances trustworthy coronary angiography analysis by employing DPO to align vision-language perception with topological constraints and utilizing RL-based reasoning with an explicit rejection mechanism, thereby achieving state-of-the-art vessel segmentation and significantly reducing false positives while prioritizing diagnostic reliability over mere coverage.

The Gist

Imagine you are trying to navigate a complex, foggy maze made of tiny, winding rivers. Your goal is to find the narrowest, most dangerous parts of the rivers (stenosis) where a boat might get stuck.

For a long time, computers trying to solve this maze had two big problems:

1. The "Pixel Puzzle" Problem: They were great at coloring in the river pixels correctly, but they kept drawing the river as a bunch of disconnected islands. It looked like a river, but if you tried to sail it, you'd fall off the edge because the computer didn't understand that rivers must be connected.
2. The "False Alarm" Problem: When the computer saw a fork in the river or a place where two rivers crossed, it would panic and scream, "Danger! Narrowing!" when it was actually just a normal part of the river's shape. This made doctors tired of ignoring the computer's alarms.

Enter ARIADNE.

Named after the Greek figure who gave Theseus a thread to navigate the Labyrinth, this new AI framework is designed to be the ultimate guide for doctors looking at heart artery images. It solves the maze using a two-step team approach: a Perceptionist and a Reasoner.

Step 1: The Perceptionist (The "Thread" Keeper)

The Job: Drawing the map.

Traditional AI tries to win a game by matching every single colored dot (pixel) perfectly. But in medicine, a perfect dot-match isn't enough if the river is broken.

ARIADNE's Perceptionist uses a special trick called DPO (Direct Preference Optimization). Think of this like teaching a student not just by giving them a test, but by showing them two drawings and saying:

• "This drawing has a broken river. Bad."
• "This drawing has a continuous, flowing river, even if the edges are slightly fuzzy. Good."

By constantly showing the AI examples of "connected" vs. "broken" rivers, it learns a new rule: "Continuity is more important than perfect edges." It learns to hold the "thread" of the river together, ensuring the map it draws is a single, unbroken path, just like a real blood vessel.

Step 2: The Reasoner (The "Smart Navigator")

The Job: Finding the danger spots.

Once the Perceptionist has drawn a perfect, connected map, the Reasoner steps in. This is a Reinforcement Learning agent, which is like a video game character learning to play by trial and error.

The Reasoner walks along the river map looking for narrow spots. But here is the genius part: It has a "Skip" button.

• Old AI: Sees a fork in the river, thinks "Narrowing!" and raises a false alarm.
• ARIADNE's Reasoner: Sees a fork, thinks, "Hmm, this looks like a tricky fork, not a blockage. I'm not 100% sure. I will skip this and ask a human doctor to check it later."

This "Skip" button is crucial. Instead of trying to guess everything and making lots of mistakes, the AI admits when it's confused. This shifts the goal from "finding as many things as possible" to "finding only the things we are sure about."

The Result: A Trustworthy Co-Pilot

When the team works together:

1. The Perceptionist draws a map where the rivers never break.
2. The Reasoner walks that map, ignoring confusing forks and only flagging the real blockages.

Why does this matter?

• Fewer False Alarms: The system stops crying wolf. It reduced false alarms by 41%, meaning doctors won't be annoyed by constant, unnecessary alerts.
• Better Safety: It found the real blockages just as well as the best existing systems, but with much higher confidence.
• Generalization: It works well even when looking at images from different hospitals or with different camera settings, proving it learned the rules of anatomy, not just memorized the pictures.

The Big Picture

This paper is a breakthrough because it realizes that for medical AI, being "smart" isn't just about seeing details; it's about understanding structure.

Just as you wouldn't trust a GPS that draws a road that disappears into a cliff, you can't trust a medical AI that draws a blood vessel that breaks in half. ARIADNE teaches the AI to respect the "thread" of life, making it a reliable partner for doctors in saving lives.

Technical Summary

1. Problem Statement

The paper addresses two critical limitations in current automated coronary angiography analysis:

• The Semantic-Topological Gap: Conventional segmentation models (CNNs, ViTs, and Foundation Models like MedSAM) optimize for pixel-level accuracy (e.g., Dice score) but fail to enforce topological continuity. This results in fragmented vascular trees where distal branches are disconnected, rendering the output useless for hemodynamic analysis or centerline extraction.
• High False Positive Rates in Stenosis Detection: Current detection systems rely on deterministic geometric algorithms or generic object detectors. These struggle to distinguish pathological stenosis from anatomical artifacts (e.g., vessel bifurcations, crossings, and foreshortening), leading to high false positive rates and "alert fatigue" for clinicians.
• Lack of Clinical Reasoning: Existing pipelines treat segmentation and detection as independent tasks, lacking the hierarchical reasoning process used by expert cardiologists to defer ambiguous cases.

2. Methodology: The ARIADNE Framework

ARIADNE (Anatomy-aware Reasoning for Integrated Angiography Diagnosis and Navigation Expert) is a two-stage framework coupling Preference-Aligned Perception with Reinforcement Learning (RL) based Reasoning.

Stage 1: Anatomy-Aware Perception (Topological Alignment)

This stage utilizes the Sa2VA (Segment Anything 2 + Vision-Language) foundation model, fine-tuned to prioritize structural continuity over pixel overlap.

• Direct Preference Optimization (DPO): Instead of standard supervised learning, the authors use DPO to align the model with topological constraints.
  • Preference Construction: A preference dataset is created where "winning" samples are topologically valid masks (Betti number $\beta_0 = 1$, indicating a single connected component) and "losing" samples are fragmented masks that may still have high pixel-level Dice scores.
  • Objective: The model learns to maximize the likelihood margin between valid and invalid topologies, effectively learning that a 92% Dice score with connectivity is superior to a 95% Dice score with fragmentation.
• Hard Sample Focused Training (HSFT): To handle anatomically challenging regions (distal vessels, bifurcations), the framework identifies "temporal hard clusters" (frames with low confidence) and applies a hybrid loss function ($L_{Dice} + \lambda L_{BCE}$) to refine boundaries in these specific areas without wasting resources on easy samples.
• Physiological Sampling: The training leverages video sequences, sampling frames across the cardiac cycle (systole/diastole) and contrast phases to ensure robustness against motion and varying contrast.

Stage 2: Structure-Guided Reasoning (Stenosis Detection)

This stage treats stenosis detection as a sequential decision-making problem modeled as a Markov Decision Process (MDP).

• Input: The agent navigates the vessel centerline extracted from the topologically consistent segmentation mask.
• State Space: Encodes local geometric features including radius profiles, first/second derivatives (gradients/curvature), Z-scores relative to baseline caliber, and local curvature.
• Action Space: Includes navigation commands (Left/Right) and diagnostic actions (Confirm or Reject).
• Explicit Rejection Mechanism: A critical innovation where the agent can abstain from making a prediction on ambiguous candidates (e.g., bifurcations that mimic narrowing). This shifts the paradigm from "maximizing coverage" to "maximizing reliability."
• Reward Function: Asymmetric rewards are used to heavily penalize False Negatives (missing a stenosis) while rewarding the correct rejection of anatomical artifacts.
• Algorithm: The agent is trained using Proximal Policy Optimization (PPO) with a Multi-Layer Perceptron (MLP) policy network.

3. Key Contributions

1. First Application of DPO for Topological Alignment: The paper introduces the use of Direct Preference Optimization in medical imaging to enforce structural constraints (vascular connectivity) rather than just semantic correctness.
2. Perception-Reasoning Synergy: It bridges the gap between visual perception and clinical reasoning by using a topologically coherent segmentation as the scaffold for an RL-based diagnostic agent.
3. Rejection-Enabled Workflow: The integration of an explicit rejection mechanism allows the system to mimic clinical triage, deferring ambiguous cases to human review, thereby significantly reducing false positives without sacrificing sensitivity.
4. Hard Sample Mining Strategy: The HSFT strategy improves computational efficiency by focusing training resources on the most diagnostically difficult anatomical regions.

4. Experimental Results

The framework was evaluated on 1,400 clinical angiograms (internal dataset) and external benchmarks (ARCADE and XCAD).

• Segmentation Performance:

  • Achieved a state-of-the-art Centerline Dice (clDice) of 0.838, significantly outperforming MedSAM3 (0.7105) and standard U-Net variants.
  • Demonstrated superior Normalized Surface Dice (NSD) and clDice, proving effective preservation of vascular connectivity even in low-contrast "wash-out" phases.
  • Showed strong generalization on the external XCAD dataset, maintaining high performance despite domain shifts.

• Stenosis Detection Performance:

  • True Positive Rate (TPR): 0.867 (outperforming baselines like Stenunet at 0.812).
  • False Positives Per Image (FPPI): Reduced to 0.85, a massive improvement over baselines (1.89–2.45), directly addressing alert fatigue.
  • F1 Score: 0.732, representing a balanced optimization of sensitivity and specificity.

• Efficiency: Inference time is <127ms per frame on a V100 GPU, suitable for real-time clinical integration.

5. Significance and Impact

• Bridging the Gap: The study proves that scaling foundation models alone is insufficient for medical imaging; explicit alignment with domain-specific anatomical priors (topology) is required.
• Clinical Trustworthiness: By integrating a rejection mechanism, the system aligns with the safety-critical nature of interventional cardiology, reducing the burden of false alarms on clinicians.
• Methodological Transfer: It successfully transfers alignment techniques from Natural Language Processing (DPO) to geometric medical image analysis, offering a new paradigm for training foundation models in safety-critical domains.
• Future Directions: The authors note that while 2D X-ray has limitations (foreshortening), this framework provides a robust foundation for future integration with 3D modalities (CTA, IVUS) and multi-view fusion.

In conclusion, ARIADNE demonstrates that anatomical validity is a prerequisite for reliable automated diagnosis, offering a pathway to transform passive image storage into active, trustworthy clinical intelligence.