From Segmentation to Structured Reports: Auditable Continual Learning for Automation in Medical Imaging Diagnostics

Le résumé fourni par la source

Autonomous medical systems such as surgical robots, AI-guided treatment planners, and telemedicine platforms require diagnostic modules that are adaptive, interpretable, and structured. We present a modular, continually learning perception-to-report pipeline for automated brain tumor diagnosis, designed for integration into clinical robotic ecosystems. The framework addresses three core requirements. (1) Adaptation: A two-stage Mixture-of-Modality-Experts (MoME+) strategy first trains four modality-specific 3D U-Net experts independently, then trains a hierarchical gating and fusion network over frozen experts. This enables selective knowledge updates without catastrophic forgetting, further supported by Elastic Weight Consolidation (EWC) and experience replay for continual learning. (2) Auditability: A deterministic registration pipeline maps segmentation outputs to the Harvard-Oxford brain atlas, producing certifiable anatomical descriptors with explicit spatial provenance. (3) Structured Human-Machine Interface: A JSON-mediated intermediate representation encodes anatomical findings with full traceability, supporting both automated downstream processing and operator-level verification, which is a key safety requirement for robotic-assisted diagnostics. Expert training on BraTS 2024 (GLI, 1,620 cases) yields per-expert Dice scores of 0.7889 (T1CE), 0.7438 (T2), 0.7206 (T1), and 0.6991 (FLAIR), with the gating fusion network achieving a mean Dice of $\mathbf{0. 8 1 5 0}$. Report generation achieves BLEU-4: 44.73, ROUGE-L: 0.653, and METEOR: 0.397 with zero hallucination. This work provides a practical architecture for certifiable diagnostic intelligence in medical robotic systems.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.