NeuroLangSeg: Language-Guided Subcortical Segmentation with Pseudo-Supervision and Anatomical-Linguistic Validation.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Vision-language models and LLMs advance contextual anatomical reasoning in brain MRI segmentation.
• A key limitation is the lack of unified anatomical definitions and consistent annotation standards in existing datasets.
• This heterogeneity hinders cross-model comparison and valid anatomical measurements.

Purpose of the Study:

• To introduce NeuroLangSeg, a language-guided framework enforcing a consistent anatomical protocol for subcortical segmentation.
• To address the limitations of heterogeneous labeling systems and enable anatomically consistent segmentations.
• To support subject-level reporting grounded in a unified anatomical standard.

Main Methods:

• NeuroLangSeg integrates a pretrained image encoder with protocol-aligned anatomical prompts and a masked pseudo-labeling strategy.
• An anatomical-linguistic evaluator acts as a training discriminator, assessing shape, spatial relationships, and volumetric norms.
• The framework enables data-efficient and interpretable learning under limited supervision.

Main Results:

• NeuroLangSeg achieves improved segmentation accuracy, with +4.1 DSC / +8.0 NSD in in-site settings and +3.6 DSC / +14.5 NSD in cross-site generalization.
• The LLM-visual integration enables anatomically verifiable predictions.
• The framework yields consistent segmentations suitable for research and clinical use.

Conclusions:

• NeuroLangSeg provides a unified anatomical standard for subcortical segmentation in brain MRI.
• The language-guided framework enhances segmentation consistency and accuracy.
• This approach facilitates reliable anatomical measurements and reporting across diverse datasets.