Evaluating the Diagnostic Classification Ability of Multimodal Large Language Models: Insights from the Osteoarthritis Initiative

2026-01-05

arXiv.org

AI Summary

This study investigates the underwhelming performance of multimodal large language models (MLLMs) on knee osteoarthritis classification from X-ray images. Through systematic ablation studies, the authors evaluate the contributions of individual components—including visual encoders, connector modules, and large language models—and optimize the system via LoRA fine-tuning and prompt engineering. Their findings reveal that fine-tuning only the visual encoder can outperform the full MLLM pipeline. Moreover, training on a small-scale yet balanced, high-quality dataset yields significantly better results than leveraging large-scale but imbalanced data. These results suggest that, for this specific medical imaging task, MLLMs are better suited as interpreters rather than primary classifiers, underscoring that data quality and class balance are more critical than sheer dataset size.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Multi-modal Vision
• Data Mining & Knowledge Management: Mining of Visual, Multimedia & Multimodal Data

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Multimodal large language models (MLLMs) show promising performance on medical visual question answering (VQA) and report generation, but these generation and explanation abilities do not reliably transfer to disease-specific classification. We evaluated MLLM architectures on knee osteoarthritis (OA) radiograph classification, which remains underrepresented in existing medical MLLM benchmarks, even though knee OA affects an estimated 300 to 400 million people worldwide. Through systematic ablation studies manipulating the vision encoder, the connector, and the large language model (LLM) across diverse training strategies, we measured each component's contribution to diagnostic accuracy. In our classification task, a trained vision encoder alone could outperform full MLLM pipelines in classification accuracy and fine-tuning the LLM provided no meaningful improvement over prompt-based guidance. And LoRA fine-tuning on a small, class-balanced dataset (500 images) gave better results than training on a much larger but class-imbalanced set (5,778 images), indicating that data balance and quality can matter more than raw scale for this task. These findings suggest that for domain-specific medical classification, LLMs are more effective as interpreters and report generators rather than as primary classifiers. Therefore, the MLLM architecture appears less suitable for medical image diagnostic classification tasks that demand high certainty. We recommend prioritizing vision encoder optimization and careful dataset curation when developing clinically applicable systems.

Problem

• multimodal large language models
• diagnostic classification
• knee osteoarthritis
• medical image classification
• disease-specific classification

Innovation

• multimodal large language models
• medical image classification
• ablation study
• LoRA fine-tuning
• data balance