Beyond Discrete Tokens: Preserving Numerical Continuity in Medical Vision-Language Models

TL;DR

HAD-VLM introduces a hybrid autoregressive-diffusion framework that unifies language generation with continuous numerical prediction for medical imaging, overcoming the limitations of discrete token-based approaches. It achieves state-of-the-art accuracy in echocardiographic measurements and improves vertebral length estimation by leveraging diffusion-based coordinate-length prediction and consistency post-processing.

Why It Matters

This work enables more accurate, reliable quantitative measurements in clinical settings by preserving numerical continuity, which is critical for diagnosis and treatment planning. The unified framework supports both precise measurements and natural language interaction, enhancing usability in real-world medical workflows.

Key Findings

• HAD-VLM reduces mean absolute error by 21.4-37.6% compared to autoregressive baselines on echocardiographic measurements.
• The diffusion-based continuous prediction with coordinate-length consistency improves measurement accuracy and robustness.
• The model maintains strong performance on medical question answering and generalizes to X-ray vertebral measurements with a 41.2% error reduction.