ISRNet: Inter-Slice Refinement in a U-Net architecture for robust WMH segmentation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

White matter hyperintensities (WMH) are important imaging biomarkers of cerebral small vessel disease and are closely associated with cognitive decline. However, the inherent anisotropy of clinical MRI remains challenging for existing segmentation methods. Standard 2D approaches neglect inter-slice anatomical continuity, whereas the effectiveness of 3D convolutions is often compromised by large slice gaps, which can introduce noise and artifacts. The central problem addressed in this study is how to preserve volumetric continuity for WMH segmentation in anisotropic MRI without forcing unreliable feature aggregation across distant slices. To address this problem, we propose ISRNet, a parameter-efficient framework that adapts the Segment Anything Model 2 (SAM2) for volumetric WMH segmentation. Unlike conventional 2D methods that process slices independently, rigid 3D methods that aggregate neighboring slices with fixed kernels, and existing SAM-based adaptations that mainly remain slice-wise, ISRNet explicitly refines the current slice with anatomically relevant context from bidirectional adjacent slices and the prior predicted mask. This refinement is implemented by the proposed Inter-Slice Refinement (ISR) module, which uses attention-based selective aggregation to combine the flexibility of 2D representation learning with the volumetric consistency sought in 3D segmentation. We evaluated ISRNet on the public MICCAI 2017 WMH Challenge dataset and a challenging private low-field (0.35T) dataset with large slice thickness (7.5 mm). ISRNet achieved a state-of-the-art (SOTA) Dice score of 76.2% on the public dataset, outperforming existing 2D, 3D, and SAM-based models. Notably, with only 3.5M trainable parameters, ISRNet demonstrated strong performance on the low-field dataset, achieving the highest Dice score and the lowest AVD among all evaluated methods. These results indicate that ISRNet improves lesion detection sensitivity and boundary accuracy while preserving volumetric consistency, highlighting the novelty of selectively modeling inter-slice continuity within a parameter-efficient SAM2 adaptation and its potential for reliable clinical deployment. Our implementation is available at https://github.com/Adventureeee/ISRNet .