NEUROTRUST: A CLINICALLY GROUNDED MULTI-AGENT FRAMEWORK FOR TRUSTWORTHY BRAIN TUMOR DIAGNOSIS AND TREATMENT PLANNING

Abstract

Current AI-based systems have shown significant promise for enhancing diagnosis and treatment planning for brain tumors, but they are not always easy to interpret, well integrated with clinical guidelines, or trustworthy for clinical decision support. We introduce NeuroTrust, a clinically inspired multiple-agent system, to overcome these challenges by combining six specialized artificial intelligence (AI) agents to orchestrate collaborative reasoning for brain tumor diagnosis and treatment planning. It integrates multimodal magnetic resonance imaging (MRI) methods, clinical history, verification of evidence-based guideline documents, uncertainty quantification, explainable decision-making, and human-in-the-loop validation processes to improve the reliability and clinical applicability of the framework. NeuroTrust was assessed with the BraTS 2023 challenge set as well as a separate de-identified clinical set on which it was externally validated. Experimental results show good segmentation performance with a mean Dice similarity score of 0.913 and classification accuracy of 94.28% while conforming to the clinical guideline with 95.04% accuracy. We additionally present two complementary indices of trustworthiness, Clinical Trust Score (CTS) and Uncertainty Calibration Index (UCI), that are derived to measure AI clinical decision reliability and calibration. NeuroTrust is an improvement on diagnostic performance, a significant reduction in hallucination rates, and better clinician trust arising from evidence-based reasoning and consensus-based decision-making when compared to a representative single agent and state-of-the-art basic approaches. The results indicate that multi-agent clinical intelligence, working together and sharing observations, can potentially serve as a viable approach for creating reliable AI solutions that can assist with brain tumour diagnosis and planning for clinical interventions in real-world scenarios.