MOSAIC: Multimodal OA Staging with Adaptive Interpretable Clinical Constraints for Evidence-Grounded Knee MRI Grading

Basit Ali, Sheheryar Khan, Rizwan Qureshi, Muhammad Asim, Mehmood Nawaz, Basim Azam

Keywords: Knee Osteoarthritis, Vision-Language Models, Chain-of-Thought Reasoning, Retrieval-Augmented Generation, Evidence-Grounded Reasoning

TL;DR: MOSAIC is an evidence-based knee MRI framework that replaces black-box OA grading with four-stage biomechanical reasoning, producing MOAKS-consistent grades supported by compartment-level measurements.

Abstract: Quantitative knee MRI grading for osteoarthritis (OA) remains labor-intensive, requiring expert radiologists to integrate compartment-specific tissue findings under structured rubrics such as MOAKS and WORMS. Existing automated systems either predict discrete severity grades without measurement-backed justification or rely solely on report text without direct image analysis, leaving a critical gap between algorithmic output and clinical utility. We present MOSAIC, a clinically aligned, evidence-grounded framework that addresses this gap through biomechanical reasoning. Its core components include a Biomechanical-Chain Guided Adaptive Gating Module, which enforces physiological coupling constraints between cartilage, bone, and meniscus, and a Contextual Evidence Retrieval system with LLM-RAG, which drives a structured four-stage Chain-of-Thought process for MOAKS-consistent and interpretable report generation. Region-wise imaging evidence is provided by three task-specific expert modules: an Anatomical Segmentation Expert (ASE), a MOAKS-Aligned Radiomics Expert (MARE), and a Visual Phenotype Expert (VPE). We further introduce the MOSAIC-Fidelity Index, a principled rubric for evaluating clinical reasoning quality beyond standard classification metrics. Preliminary evaluation across multi-sequence knee MRI demonstrates the feasibility of this reasoning-centered approach for robust, grading-aligned OA assessment with measurement-backed clinical justification.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 25