Deep-Learning Atlas Registration for Melanoma Brain Metastases: Preserving Pathology While Enabling Cohort-Level Analyses

1. Problem Statement

Melanoma brain metastases (MBM) are heterogeneous lesions that pose significant challenges for cohort-level analysis due to:

• Anatomical Variability: Differences in brain structure across patients.
• Geometric Distortions: Variations in MRI acquisition parameters and scanner types.
• Missing Correspondences: The fundamental challenge in atlas registration where a pathological lesion (tumor) exists in the patient image but has no corresponding region in the healthy anatomical atlas.
  • Consequence: Standard deformable registration algorithms often collapse or distort the tumor volume to match the healthy atlas, destroying pathological information and rendering spatial analysis inaccurate.
  • Current Limitations: Existing solutions often require extensive preprocessing (e.g., skull stripping, intensity normalization), manual lesion masking, or tuned thresholds to handle missing correspondences, limiting their clinical adoption and reproducibility.

2. Methodology

The authors propose a fully differentiable, preprocessing-free, deep-learning-based deformable registration framework.

A. Core Architecture

• Sampling Module: Instead of resampling and interpolating images during a preprocessing step (which degrades image quality), the framework uses a "sampling module" wrapper. This module stores a sampling grid for each image. Transformations are applied to the grid, not the image itself, preserving the original image data and minimizing interpolation artifacts.
• Neural Network: A U-Net architecture is used to predict a stationary velocity field ($v$).
  • Input: Patient image ($I_i$) and Atlas image ($A$).
  • Output: Velocity field integrated via a "scaling and squaring" layer to generate the final diffeomorphic transformation ($T_i$) and its inverse ($T_i^{-1}$).
  • Connections: Uses skip connections via summation rather than concatenation.

B. Training Strategy (Two-Step Approach)

1. General Model (Forward Approach): Trained on a large dataset (Learn2Reg) to learn a baseline registration.
   • Loss Function ($L_G$): Includes similarity loss ($L_{sim}$) between the transformed atlas and the patient image in image space, a regularization term ($L_{reg}$), and a cross-dataset similarity term.
   • Key Insight: Evaluating similarity in image space inherently penalizes the collapse of the tumor because shrinking the tumor reduces the overlap with the atlas in the patient's native space.
2. One-Shot Over-fitting (Backward Approach): The general model is fine-tuned (over-fitted) for each specific patient case to maximize accuracy.
   • Loss Function ($L_{OF}$): Includes similarity loss in atlas space, regularization, and a Volume-Preserving Loss ($L_{vol}$).
   • Volume Preservation: This term ensures that the volume change of the tumor region is similar to the surrounding healthy tissue, preventing unrealistic collapse without needing a pre-defined tumor mask.

C. Handling Missing Correspondences

The framework addresses the "missing correspondence" problem (tumor vs. healthy atlas) without explicit lesion masks:

• Distance-Transformed Labels: Instead of comparing raw voxel intensities, the method converts anatomical label maps into distance maps. The similarity loss is calculated as the negative Normalized Cross-Correlation (NCC) between these distance maps.
• Mechanism: By evaluating similarity in image space (during general training) and enforcing volume preservation (during over-fitting), the network is discouraged from collapsing the tumor to match the healthy atlas.

D. Datasets

• Training/Validation: 414 healthy subjects from the 2021 Learn2Reg challenge.
• Clinical Cohort: 209 patients with MBM from three independent German centers (Clin1–3).
• Atlases Used: ICBM152 (anatomical), Arterial Territory Atlas (32 territories), and Perfusion Atlas.

3. Key Contributions

1. Preprocessing-Free Registration: The framework operates directly on native medical images, eliminating the need for skull stripping, intensity normalization, or resampling, thereby preserving image fidelity.
2. Mask-Free Pathology Handling: It robustly handles missing correspondences (tumors) without requiring manual lesion masks or threshold tuning, relying instead on forward-model similarity evaluation and volume-preserving losses.
3. Publicly Available Tool: The complete code and pre-trained models are open-sourced, facilitating reproducible research.
4. Clinical Application: Successfully applied to a multi-center MBM cohort to perform standardized spatial analysis.

4. Results

A. Registration Accuracy

The framework demonstrated high accuracy across all datasets:

• Dice Similarity Coefficient (DSC):
  • Learn2Reg (Over-fitted): 0.92 ± 0.09
  • Clinical Datasets (Clin1–3, Over-fitted): 0.89 – 0.90
• Hausdorff Distance (HD): Reduced to 6.79 – 7.60 mm.
• Average Symmetric Surface Distance (ASSD): Reduced to 0.63 – 0.77 mm.
• Tumor Preservation:
  • Without volume-preserving loss, tumor volume decreased by a factor of 0.73.
  • With volume-preserving loss, tumor volume decreased by only 0.84 (statistically significant improvement).
  • The Jacobian determinant ratio (tumor vs. 10mm peritumoral region) remained close to 1.0 (0.93 ± 0.37), indicating anatomically plausible deformations.

B. Clinical Spatial Analysis (MBM Distribution)

Applying the framework to 209 patients revealed:

• Gray-White Matter Junction: Metastases were significantly concentrated near the gray-white matter junction ($p < 0.001$), confirming a known biological predilection.
• Cortical vs. White Matter:
  • Over-represented: Cerebral Cortex (309 observed vs. 224 expected) and Putamen (20 observed vs. 5 expected).
  • Under-represented: Cerebral White Matter (131 observed vs. 194 expected).
• Arterial Territories: No single arterial territory showed a significantly increased frequency of metastases after correcting for regional volume.
• Corticomeningeal Interface: A secondary preference for the interface between the cortex and meninges was observed.

5. Significance and Conclusion

• Scientific Impact: The study validates and refines existing knowledge regarding MBM spatial distribution, confirming the preference for the gray-white matter junction and cortical regions while challenging previous assumptions about arterial territory dominance.
• Methodological Advancement: By removing the dependency on lesion masks and extensive preprocessing, this framework lowers the barrier for large-scale, multi-center neuro-oncological studies. It enables the creation of "digital twins" of pathological brains aligned to a common atlas.
• Future Utility: The tool supports reproducible cohort-level analyses, radiomics, and the identification of anatomical regions associated with recurrence or treatment toxicity. It is designed to be extensible to other tumor types and pathological brain conditions.

In summary, this work presents a robust, deep-learning solution that solves the "missing correspondence" problem in atlas registration, enabling accurate, large-scale spatial analysis of brain metastases without compromising pathological data integrity.