MedSAM/MedSAM2 Feature Fusion: Enhancing nnUNet for 2D TOF-MRA Brain Vessel Segmentation

TL;DR

This paper improves brain vessel segmentation in MRI scans by combining the nnUNet framework with MedSAM/MedSAM2 features, leading to better accuracy and boundary detection.

Contribution

The novel integration of MedSAM/MedSAM2 features into nnUNet for enhanced vessel segmentation in TOF-MRA brain slices.

Findings

The MedSAM2-enhanced model improved Dice coefficient by 0.72% compared to the baseline.

HD95 and ASD were reduced to 46.30 mm and 4.97 mm, respectively, showing better boundary localization.

The method shows potential for improving cerebrovascular disease diagnosis in clinical practice.

Abstract

Accurate segmentation of brain vessels is critical for diagnosing cerebral stroke, yet existing AI-based methods struggle with challenges such as small vessel segmentation and class imbalance. To address this, our study proposes a novel 2D segmentation method based on the nnUNet framework, enhanced with MedSAM/MedSAM2 features, for arterial vessel segmentation in time-of-flight magnetic resonance angiography (TOF-MRA) brain slices. The approach first constructs a baseline segmentation network using nnUNet, then incorporates MedSAM/MedSAM2’s feature extraction module to enhance feature representation. Additionally, focal loss is introduced to address class imbalance. Experimental results on the CAS2023 dataset demonstrate that the MedSAM2-enhanced model achieves a 0.72% relative improvement in Dice coefficient and reduces HD95 (mm) and ASD (mm) from 48.20 mm to 46.30 mm and from 5.33 mm to 4.97 mm, respectively, compared to the baseline nnUNet, showing significant enhancements in boundary localization and segmentation accuracy. This approach addresses the critical challenge of small vessel segmentation in TOF-MRA, with the potential to improve cerebrovascular disease diagnosis in clinical practice.

Figures