From Coarse Prediction to Precise Delineation: Foundation Model-Guided White Matter Lesion Segmentation on T2-FLAIR MRI
Object To develop a robust segmentation framework for white matter lesions, including white matter hyperintensities (WMH) and multiple sclerosis (MS) lesions, on T2 fluid-attenuated inversion recovery (FLAIR) MRI by integrating a foundation model with a U-Net-based architecture.
Methods Two public datasets were used: the 2017 MICCAI WMH Segmentation Challenge (MWSC) dataset, comprising 170 subjects (60 training and 110 test subjects), and the MSLesSeg dataset, comprising 115 scans from 75 patients (93 training and 22 test scans). Models were trained and evaluated separately for each dataset using only FLAIR images. The proposed framework employs a two-stage segmentation strategy that integrates nnUNet with the Segment Anything Model (SAM), a widely used segmentation foundation model. SAM was fine-tuned to improve adaptation to medical image characteristics. First, nnUNet generates a coarse segmentation that captures the overall lesion structure. This prediction is subsequently used as a structural prompt to guide SAM in a slice-wise manner, together with the original image. By providing an explicit lesion prior, the framework constrains the refinement process while preserving SAM’s ability to delineate lesion boundaries.
Results Performance was evaluated using the Dice similarity coefficient (DSC) and absolute volume difference (AVD). The proposed method outperformed the baseline on both datasets. On the MWSC dataset, it achieved a DSC of 81.12% (95% CI: 79.41–82.83) and an AVD of 14.10% (95% CI: 11.47–16.74), corresponding to a 6.18% improvement in DSC and a 9.24% reduction in AVD (both p < 0.001). On the MSLesSeg dataset, it achieved a DSC of 72.18% (95% CI: 67.72–76.63), an AVD of 14.58% (95% CI: 8.98–20.19), improving DSC by 3.48% (p = 0.002) and reducing AVD by 11.13% (p = 0.025).
Conclusions Foundation model-guided refinement with U-Net-derived coarse structural priors improves segmentation of WMH and MS lesions on FLAIR MRI, enhancing lesion localization and boundary delineation and enabling more consistent lesion quantification.