Native space based pipelines outperform template space based pipeline in subcortical segmentation

Technical Summary: Native Space vs. Template Space Pipelines for Subcortical Segmentation

Problem Statement
Accurate segmentation of subcortical nuclei, specifically the Subthalamic Nucleus (STN), Red Nucleus (RN), and Substantia Nigra (SN), is critical for neurosurgical planning in Parkinson's Disease (PD) and functional research. While high-field (7T) MRI offers superior tissue contrast, clinical practice predominantly relies on lower-field (3T) scans where these structures are difficult to identify due to limited contrast. Current automated segmentation tools typically rely on template space coregistration (e.g., warping patient images to the MNI standard space) to leverage population-averaged atlases. However, this approach introduces a "back-and-forth" registration process: transforming the patient to template space for segmentation and then inverse-transforming the results back to the native space. The authors hypothesize that this normalization and inverse transformation may compromise patient-specific anatomical fidelity, particularly for small, irregular structures like the STN, and that native space approaches may offer a measurable advantage. Furthermore, the transferability of models trained on high-quality 7T data to clinical 3T data remains an open challenge, prompting an investigation into whether synthetic 3T data can bridge this domain gap.

Methodology
The study developed and compared two deep learning-based segmentation pipelines using the nnUNet framework (specifically UNet with Residual Encoder presets) trained on a curated dataset of 7T and 3T T2-weighted MRI images from five public sources (CBS, ATAG, Chu 2025, Paired, and OASIS).

1. Template Space Pipeline:

   • All subject images were affine-registered to the MNI Nonlinear Asymmetric 2009 template.
   • Two template resolutions were tested: 1 mm (MNI 152 NLIN 2009c) and 0.5 mm (MNI 152 NLIN 2009b).
   • A Region of Interest (ROI) was defined in MNI space, the model was trained on these normalized images, and the output segmentations were inverse-transformed back to the native space for evaluation.

2. Native Space Pipeline:

   • Operated directly on the patient's native MRI images.
   • Subject-specific ROIs were constructed by warping the CIT168 atlas (probabilistic maps) to the native space to define bounding boxes around the target structures.
   • The model was trained and evaluated entirely within the native space, avoiding inverse transformations of the final segmentation labels.

3. Domain Adaptation (Synthetic Data):

   • To address the performance drop when applying 7T-trained models to 3T data, the authors generated synthetic 3T images from 7T training data using two methods:
     • Simple: Histogram standardization via piecewise-linear matching.
     • Advanced: A modified MURD (Multi-site Unsupervised Representation Disentangler) framework to disentangle anatomical content from site-specific style (intensity variations) and generate 3T-like images.
   • Models trained on 7T data were tested with and without these synthetic 3T augmentations to evaluate knowledge transferability.

Key Contributions

• Comparative Evaluation: A direct, quantitative comparison of native space versus template space segmentation pipelines for subcortical structures, specifically isolating the impact of the inverse transformation step.
• Resolution Analysis: An investigation into whether increasing template resolution (0.5 mm vs. 1 mm) mitigates the accuracy loss associated with template-based workflows.
• Domain Gap Investigation: A systematic evaluation of 7T-to-3T model transferability and the efficacy of synthetic data augmentation (histogram matching vs. MURD) in bridging the field-strength gap without requiring manual 3T labels.

Results

• Native vs. Template (7T Data): On held-out 7T test data, the native space pipeline consistently outperformed the template-based approaches.
  • For the STN, the native method achieved a Dice score of 0.775 ± 0.055 compared to 0.713 ± 0.051 for the 1 mm template and 0.650 ± 0.166 for the 0.5 mm template.
  • The native method also showed superior boundary accuracy (lower HD95) for the STN and SN.
  • Counter-intuitive Finding: Increasing the template resolution from 1 mm to 0.5 mm degraded performance, suggesting that the error source is the affine transformation/interpolation process itself rather than insufficient template resolution.
• 7T-to-3T Transferability: Models trained exclusively on 7T data suffered a significant performance drop when applied to 3T images (e.g., STN Dice dropped from ~0.775 to ~0.556 for the native pipeline).
• Synthetic Data Impact:
  • Adding synthetic 3T data (generated via MURD or histogram matching) yielded only modest and inconsistent improvements for 3T segmentation performance.
  • While MURD improved the Red Nucleus (RN) Dice and HD95 in the template pipeline, it offered limited gains for the STN and SN.
  • Crucially, the addition of synthetic data did not significantly degrade performance on 7T test data, indicating that mixed-field training is safe but not necessarily transformative for 3T generalization in this context.

Significance and Claims
The paper concludes that native space segmentation is preferable for applications requiring high patient-specific anatomical fidelity, such as surgical planning in PD. The study demonstrates that the inverse affine transformation required by template-based workflows introduces a meaningful bottleneck, particularly for small, variable structures like the STN, where even high-resolution templates fail to compensate for registration errors.

Regarding domain adaptation, the authors claim that while synthetic 3T data generation is a viable strategy that does not harm 7T performance, it does not fully bridge the 7T-to-3T domain gap. The limitations of current style-transfer methods (like MURD) in replicating the complex contrast-to-noise ratios and tissue boundary visibility differences between field strengths suggest that bridging this gap remains an open challenge requiring future work on domain adaptation tailored specifically to subcortical structures. The findings support a shift toward patient-specific segmentation pipelines to avoid the accuracy loss inherent in spatial normalization.