HAD-VLM
A medical vision-language model with a hybrid autoregressive-diffusion architecture. It generates continuous numerical representations through a diffusion process and is used for quantitative measurement tasks in medical imaging.
Timeline (last 90 days)
HAD-VLM adopts a hybrid autoregressive-diffusion architecture that retains both an autoregressive text generation pathway and a diffusion-based continuous prediction pathway within a shared backbone.
HAD-VLM's diffusion head generates a five-dimensional continuous vector containing four caliper endpoint coordinates and one measurement length.
HAD-VLM proposes a coordinate-length consistency post-processing mechanism that cross-checks the distance computed from endpoint coordinates against the predicted length, combined with ensemble prediction refinement.
On the EchoNet-LVH dataset, HAD-VLM's mean absolute errors on four echocardiographic measurement tasks are: IVSd 1.36 mm, LVIDd 2.77 mm, LVIDs 3.05 mm, and LVPWd 1.16 mm.
Relative to the Vicuna autoregressive baseline, HAD-VLM reduces error by 21.4%–37.6% on the four EchoNet-LVH tasks.
HAD-VLM outperforms the fine-tuned Qwen3-VL and the direct MLP regression method on all four EchoNet-LVH measurement tasks.
In the X-ray vertebral measurement task, HAD-VLM's mean absolute error drops from 11.4 pixels to 6.7 pixels.
After introducing diffusion prediction, HAD-VLM's medical question-answering performance is preserved on the evaluation datasets.
HAD-VLM's vision side uses visual prompt tuning to adapt to ultrasound images, freezing most of the pretrained visual encoder and inserting only a small number of learnable prompt vectors.
HAD-VLM's experiments are confined to two task types, cardiac ultrasound and vertebral measurement, and its sample scale and clinical diversity remain limited.
1 more timeline events
All Facts (11)
HAD-VLM adopts a hybrid autoregressive-diffusion architecture that retains both an autoregressive text generation pathway and a diffusion-based continuous prediction pathway within a shared backbone.
50%UnverifiedHAD-VLM's diffusion head generates a five-dimensional continuous vector containing four caliper endpoint coordinates and one measurement length.
50%UnverifiedHAD-VLM proposes a coordinate-length consistency post-processing mechanism that cross-checks the distance computed from endpoint coordinates against the predicted length, combined with ensemble prediction refinement.
50%UnverifiedOn the EchoNet-LVH dataset, HAD-VLM's mean absolute errors on four echocardiographic measurement tasks are: IVSd 1.36 mm, LVIDd 2.77 mm, LVIDs 3.05 mm, and LVPWd 1.16 mm.
50%UnverifiedRelative to the Vicuna autoregressive baseline, HAD-VLM reduces error by 21.4%–37.6% on the four EchoNet-LVH tasks.
50%UnverifiedHAD-VLM outperforms the fine-tuned Qwen3-VL and the direct MLP regression method on all four EchoNet-LVH measurement tasks.
50%UnverifiedIn the X-ray vertebral measurement task, HAD-VLM's mean absolute error drops from 11.4 pixels to 6.7 pixels.
50%UnverifiedAfter introducing diffusion prediction, HAD-VLM's medical question-answering performance is preserved on the evaluation datasets.
50%UnverifiedHAD-VLM's vision side uses visual prompt tuning to adapt to ultrasound images, freezing most of the pretrained visual encoder and inserting only a small number of learnable prompt vectors.
50%UnverifiedHAD-VLM's experiments are confined to two task types, cardiac ultrasound and vertebral measurement, and its sample scale and clinical diversity remain limited.
50%UnverifiedHAD-VLM's coordinate-length consistency mechanism relies on the caliper-endpoint annotation format and may need to be redesigned when extended to free-form lesion measurement.
50%