Aspect: Explainable AI for Clinical Awareness of Autosegmentation Failure Risk
Abstract
Purpose
AI-based autosegmentation is widely used in radiation oncology to improve efficiency and consistency; however, these models may silently fail when applied to cases that deviate from their training distribution, placing responsibility on clinicians to detect unreliable results through manual review. This work presents ASPECT (Automated Segmentation Pre-check and Evaluation Context Tool), an agentic AI decision-support system designed to assist clinical users by identifying potentially unreliable auto-segmentation cases, providing interpretable explanations of contributing factors, and delivering focused review and corrective guidance.
Methods
A cohort of 1,369 prostate cases was analyzed. Outliers were defined as cases with Dice Similarity Coefficient (DSC) for bladder, prostate, or rectum below the 1st percentile. ASPECT identifies high-risk cases by analyzing multimodal features, including body habitus, clinical procedural information, imaging parameters, and auto-contour volumes. It compares patient-specific data against contextual knowledge of the training cohort, including feature distributions, representative high-quality and known failure examples. The agent performs univariate deviation assessment and multivariate reasoning to estimate case-specific failure risk and generate an interpretable QA report. The framework is implemented using LangGraph with GPT-5 (medium reasoning effort) deployed via Azure cloud services.
Results
In the test cohort, ASPECT identified 92.41% of 316 outlier cases and generated interpretable, case-specific explanations revealing the root causes of auto-segmentation failures. Incorporation of clinical notes, body habitus features, and auto-contour volumes increased detection sensitivity by 34.6% compared with image-only monitoring. Missed cases were primarily associated with unseen conditions, including rectal catheters, brachytherapy seeds and penile prostheses. The average computational cost was approximately $0.02 per case at inference.
Conclusion
ASPECT functions as a clinical decision-support “watchdog” for autosegmentation, increasing user awareness of potential failures, guiding focused review and corrective action, and preventing silent errors. By providing interpretable explanations and actionable guidance, ASPECT enhances safety, transparency, and scalability of AI-assisted contouring workflows.