Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

VesselSAM: Leveraging SAM for Aortic Vessel Segmentation with AtrousLoRA

Created by
  • Haebom

Author

Adnan Iltaf, Rayan Merghani Ahmed, Zhenxi Zhang, Bin Li, Shoujun Zhou

Outline

In this paper, we propose VesselSAM, an improved version of Segment Anything Model (SAM) for aortic vessel segmentation. VesselSAM improves its performance by introducing AtrousLoRA module, which integrates Atrous Attention and Low-Rank Adaptation (LoRA). Atrous Attention captures multi-scale contextual information to preserve both fine-grained local information and broad global information, and LoRA enables efficient fine-tuning of the frozen SAM image encoder, which reduces the number of learnable parameters and improves computational efficiency. Evaluation results using Aortic Vessel Tree (AVT) and Type-B Aortic Dissection (TBAD) datasets show that VesselSAM achieves state-of-the-art performance (DSC score above 93%) while significantly reducing the computational overhead compared to existing large-scale models.

Takeaways, Limitations

•
Takeaways:
◦
Proposal of VesselSAM, which significantly improves the performance of aortic vessel segmentation based on SAM.
◦
Leveraging multi-scale information and increasing computational efficiency with the AtrousLoRA module.
◦
Achieving state-of-the-art performance on diverse datasets.
◦
Contributing to improving AI-based aortic vessel segmentation in clinical settings.
•
Limitations:
◦
Generalization performance needs to be verified on datasets other than the presented dataset.
◦
More detailed analysis is needed to improve the performance of the AtrousLoRA module.
◦
Additional validation and safety evaluation are needed for actual clinical application.
👍
Made with Slashpage