Bootstrapping surgeon labeling campaigns with x0-diffusion: label-efficient vessel segmentation of catheter-based angiograms

Objective: Assess which architecture is the most label-efficient for vessel segmentation of catheter-based angiograms, and whether a diffusion model trained with extreme data scarcity could serve as a scaffold for expert annotation. Materials and Methods: x0-diffusion, SAM, DeepLabV3, ResNet-UNet, ViT-DPT, and ResNet-ViT hybrid were compared at 16, 32, and 80 labeled images from the 500-image Dr-SAM dataset under 5x5-fold repeated cross-validation, using one-sided paired Wilcoxon signed-rank tests with Holm correction. A vascular surgeon's annotation time, from scratch and when correcting x0-diffusion predictions, was also measured. Results: At 16 training images, x0-diffusion achieved the highest Dice (0.882) and IoU (0.795) of the six models, outperforming every competitor in all 25 paired runs on Dice, by 0.047 over zero-shot SAM and 0.226 over DeepLabV3. The advantage narrowed as labels increased: at 32 images x0-diffusion was not significantly different from the two strongest convolutional models, and at 80 images those models scored marginally higher (0.923 and 0.920 vs. 0.910). Correcting 16-image x0-diffusion predictions on five test angiograms averaged 11.1 surgeon-minutes per image versus 20.6 from scratch. Discussion: The advantage of x0-diffusion is label efficiency, not peak accuracy, with the strongest convolu- tional models needing twice the labels to match it. It rarely fails where zero-shot SAM fails, so at the start of an annotation campaign, correcting its predicted masks could cut surgeon annotation time by an estimated 37 to 45%. Conclusion: With very few labels, x0-parameterized diffusion is the strongest vessel segmenter tested and a candidate scaffold for expert annotation.

Authors

Institutions

Publication Details

Journal
medRxiv
Published
2026-11-15
DOI
https://doi.org/10.64898/2026.11.13.26368001
Primary Topic
Medical Image Segmentation Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Bootstrapping surgeon labeling campaigns with x0-diffusion: label-efficient vessel segmentation of catheter-based angiograms

Andrew Alexander Gonzalez, Shantanu Dev, Neelay Ranjan
medRxiv
Medical Image Segmentation Techniques
preprint

Bootstrapping surgeon labeling campaigns with x0-diffusion: label-efficient vessel segmentation of catheter-based angiograms

Andrew Alexander Gonzalez, Shantanu Dev, Neelay Ranjan
preprint en

Abstract

Objective: Assess which architecture is the most label-efficient for vessel segmentation of catheter-based angiograms, and whether a diffusion model trained with extreme data scarcity could serve as a scaffold for expert annotation. Materials and Methods: x0-diffusion, SAM, DeepLabV3, ResNet-UNet, ViT-DPT, and ResNet-ViT hybrid were compared at 16, 32, and 80 labeled images from the 500-image Dr-SAM dataset under 5x5-fold repeated cross-validation, using one-sided paired Wilcoxon signed-rank tests with Holm correction. A vascular surgeon's annotation time, from scratch and when correcting x0-diffusion predictions, was also measured. Results: At 16 training images, x0-diffusion achieved the highest Dice (0.882) and IoU (0.795) of the six models, outperforming every competitor in all 25 paired runs on Dice, by 0.047 over zero-shot SAM and 0.226 over DeepLabV3. The advantage narrowed as labels increased: at 32 images x0-diffusion was not significantly different from the two strongest convolutional models, and at 80 images those models scored marginally higher (0.923 and 0.920 vs. 0.910). Correcting 16-image x0-diffusion predictions on five test angiograms averaged 11.1 surgeon-minutes per image versus 20.6 from scratch. Discussion: The advantage of x0-diffusion is label efficiency, not peak accuracy, with the strongest convolu- tional models needing twice the labels to match it. It rarely fails where zero-shot SAM fails, so at the start of an annotation campaign, correcting its predicted masks could cut surgeon annotation time by an estimated 37 to 45%. Conclusion: With very few labels, x0-parameterized diffusion is the strongest vessel segmenter tested and a candidate scaffold for expert annotation.

medRxiv
Purdue University West Lafayette (US), Indiana University School of Medicine (US), The Ohio State University (US)
Medical Image Segmentation Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon