VesselSAM: a parameter-efficient adaptation of SAM to aortic vessel segmentation with AtrousLoRA

Accurate segmentation of the aorta and its branches in computed tomography angiography requires delineating fine vessels and irregular boundaries across a wide range of vessel calibers. The Segment Anything Model (SAM) provides strong pre-trained representations, but full encoder fine-tuning updates many parameters, whereas conventional Low-Rank Adaptation (LoRA) does not aggregate image context across different receptive fields within its task-specific update. We present VesselSAM, a parameter-efficient adaptation of SAM initialized from MedSAM for binary aortic vessel segmentation. AtrousLoRA adapts the query and value projections in each image-encoder block of the base vision transformer (ViT-B) by inserting a multi-branch Atrous Attention Module into each rank-4 residual path. We evaluated the model on patient-level partitions of the multi-center Aortic Vessel Tree dataset and the Type-B Aortic Dissection dataset using tight ground-truth-derived box prompts at validation and test time. VesselSAM achieved slice-averaged Dice similarity coefficient (DSC) values of 93.50%, 93.25%, and 93.02% on the Dongyang, Rider, and KiTS cohorts of the Aortic Vessel Tree dataset, respectively, and 93.26% on the Type-B Aortic Dissection dataset. AtrousLoRA achieved DSC values 1.14 to 4.50 percentage points higher than standard LoRA across the four cohorts and two initializations. VesselSAM updated 6.8 million parameters, corresponding to 7.2% of the model parameters. AtrousLoRA integrates multi-scale spatial processing into the low-rank adaptation of SAM for aortic vessel segmentation. The evaluation with ground-truth-derived box prompts showed higher DSC than standard LoRA under both SAM and MedSAM initialization.