Unlocking 2D Promptable Foundation Models for 3D Vessel Segmentation by Automatic Prompt Generation

Ziyu Zhang, Yi Yu, Yuan Xue

Keywords: Vessel Segmentation, Foundation Models, Automatic Prompting, Topology Preservation

TL;DR: We unlock 2D promptable foundation models for 3D vessel segmentation: from a single click, a topology aware framework auto generates prompts through the volume, reaching near nnUNet accuracy while cutting interaction from hours to seconds.

Abstract:

3D vessel segmentation is a core task in medical image analysis, playing a crucial role in disease diagnosis and surgical planning.
While fully supervised 3D segmentation methods rely on costly high-quality annotations, prompt-guided models (e.g., ScribblePrompt) provide a promising alternative with their zero-shot generalization capability for efficient 3D segmentation.
Nevertheless, when directly applied to 3D tasks, these 2D methods require slice-wise prompts, disregarding the continuity of 3D structures and leading to low efficiency.
To address this issue, we propose an innovative method based on automatic prompt generation, which integrates with pre-trained 2D interactive models to achieve efficient 3D vessel segmentation.
By leveraging spatial continuity and contextual information, our method automatically generates prompts across the entire 3D volume from a single user-provided prompt.
Experiments conducted on public and in-house vessel datasets demonstrate the effectiveness of the proposed method, showing that it achieves segmentation accuracy comparable to or better than state-of-the-art models, while significantly reducing the interaction cost.

Primary Subject Area: Segmentation

Secondary Subject Area: Foundation Models

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 355