TrustEndo: A Conformal-Calibrated MLLM With Retrieval-Augmented Reasoning for Trustworthy Gastroscopic Diagnosis.
TL;DR
TrustEndo is introduced, a framework that addresses both problems of hallucination of plausible-yet-wrong claims, and the lack of statistical guarantees on output reliability through three modules that extends conformal prediction to multimodal outputs, providing coverage guarantees over detection predictions under exchangeability.
Abstract
Multimodal Large Language Models (MLLMs) can generate natural language diagnostic descriptions from gastroscopic images, but their clinical use is blocked by two problems: hallucination of plausible-yet-wrong claims, and the lack of statistical guarantees on output reliability. We introduce TrustEndo, a framework that addresses both problems through three modules: (1) a Concept-Anchored Visual-Language Aligner (CAVLA) that grounds MLLM reasoning in clinically verifiable morphological concepts via text-based prompts; (2) an Evidence-Augmented Diagnostic Memory (EADM) that retrieves relevant historical cases via multi-modal RAG; and (3) a Conformal Safety Layer (CSL) that extends conformal prediction to multimodal outputs, providing coverage guarantees over detection predictions under exchangeability, with semantic-level reliability assessment for diagnostic language. Built on Qwen-3.5 and GLM-5.0, TrustEndo also includes a Domain-Adaptive Refinement (DAR) module for cross-center generalization. On LGLDD, Endo21, and HyperKvasir, TrustEndo achieves AP 40.9% (vs. 39.5% best MLLM baseline), hallucination rate 7.8% (vs. 15-20%), and ECE 0.038 (vs. 0.118+), with 95% empirical coverage on detection.