The Problem
Standard Vision-Language Models (LLaVA, MedGemma, Qwen-VL) produce fluent but clinically untrustworthy radiology reports. In neuro-oncology, misidentifying tumor grade or overlooking ependymal contact alters surgical pathways - a catastrophic failure mode standard end-to-end VLMs cannot prevent.
What I Did & Decided
Engineered ARIA, a neuro-symbolic pipeline that decouples visual extraction from language generation. SegResNet segments mpMRI into sub-regions, extracted radiomic features populate a diagnostic knowledge graph derived from official grading standards, and an LLM writes the report strictly grounded on graph attributes rather than unconstrained pixel tokens.
What Came of It (So Far)
Evaluated on 1,251 patients, segmentation reached Dice scores of WT: 0.757 [?]Whole Tumor TC: 0.428 [?]Tumor Core ET: 0.354 [?]Enhancing Tumor Upstream mask noise proved to directly alter downstream focality classification. The knowledge graph provided necessary grounding: generation scored GREEN: 0.718 while removing the graph caused hallucination rates to spike up to 0.63 across tested models.
