HAD-VLM pairs diffusion with language for ultrasound measurements
Original title: Beyond Discrete Tokens: Preserving Numerical Continuity in Medical Vision-Language Models
It cuts four EchoNet-LVH measurement errors 21.4–37.6% versus a Vicuna baseline and lowers X-ray vertebral error from 11.4 to 6.7 pixels.
Researchers present HAD-VLM, a shared-backbone model that generates language autoregressively while a diffusion process predicts continuous caliper coordinates and measurement length. Visual prompt tuning adapts the vision encoder to ultrasound. On four echocardiographic tasks it beats fine-tuned Qwen3-VL and direct MLP regression, and medical question answering on the evaluated dataset is preserved. A consistency check plus ensemble post-processing refines the estimates.
- Mean absolute errors are 1.36, 2.77, 3.05, and 1.16 mm for IVSd, LVIDd, LVIDs, and LVPWd.
- The diffusion target is a five-component representation of endpoints and length.
- Evaluation also covers medical question answering and X-ray vertebral measurement.
From the original article
1 min read at arxiv.orgObjective: To develop a unified vision-language model for prompt-conditioned quantitative ultrasound measurement that addresses the mismatch between continuous measurement targets and discrete token-based numerical prediction. Methods: This article presents HAD-VLM, a hybrid autoregressive-diffusion framework that combines autoregressive text generation with continuous measurement prediction within a shared backbone.
By Dianye Huang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.