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

TL;DR

This work tackles the variability and heterogeneity of MBM MRI by introducing a fully differentiable deformable atlas-registration framework that preserves metastatic tissue without lesion masks. It combines a sampling-module approach with a U-net-based velocity-field predictor, trained in two stages (general then one-shot over-fitting) and guided by distance-transform label maps and a volume-preserving loss to handle missing correspondences. Across seven datasets including three MBM cohorts (n=209) and multiple atlases, the method achieves high atlas-space registration accuracy (DSC around $0.89$–$0.92$) and reveals robust spatial predilections for MBM, notably near the gray-white matter junction and cortex, with no arterial-territory bias after volume correction. The framework is open-source, enabling reproducible multi-centre analyses and broad extension to other brain pathologies, thereby advancing cohort-level neuro-oncological imaging studies.

Abstract

Melanoma brain metastases (MBM) are common and spatially heterogeneous lesions, complicating cohort-level analyses due to anatomical variability and differing MRI protocols. We propose a fully differentiable, deep-learning-based deformable registration framework that aligns individual pathological brains to a common atlas while preserving metastatic tissue without requiring lesion masks or preprocessing. Missing anatomical correspondences caused by metastases are handled through a forward-model similarity metric based on distance-transformed anatomical labels, combined with a volume-preserving regularization term to ensure deformation plausibility. Registration performance was evaluated using Dice coefficient (DSC), Hausdorff distance (HD), average symmetric surface distance (ASSD), and Jacobian-based measures. The method was applied to 209 MBM patients from three centres, enabling standardized mapping of metastases to anatomical, arterial, and perfusion atlases. The framework achieved high registration accuracy across datasets (DSC 0.89-0.92, HD 6.79-7.60 mm, ASSD 0.63-0.77 mm) while preserving metastatic volumes. Spatial analysis demonstrated significant over-representation of MBM in the cerebral cortex and putamen, under-representation in white matter, and consistent localization near the gray-white matter junction. No arterial territory showed increased metastasis frequency after volume correction. This approach enables robust atlas registration of pathological brain MRI without lesion masks and supports reproducible multi-centre analyses. Applied to MBM, it confirms and refines known spatial predilections, particularly preferential seeding near the gray-white matter junction and cortical regions. The publicly available implementation facilitates reproducible research and extension to other brain tumours and neurological pathologies.

Figures (5)

• Figure 1: The overall architecture used in this work. The sampling modules act like a wrapper around images ($I_i$ and $I_j$) and the atlas ($I_A$) and are responsible for preparing them as input for the neural network and for applying the calculated transformations ($T_i$, $T_j$) and their inverse transformations ($T_i^{-1}$, $T_j^{-1}$). The neural network (here a U-net) calculates a stationary velocity field $v$, which is integrated into the final transformations $T_i$ and $T_i^{-1}$ by scaling and squaring. Blue arrows indicate data flow, while red and green arrows show which images are used for the similarity loss $L_{sim}$ of the general and the over-fitted model, respectively. Regularization losses are not shown in this figure.
• Figure 2: This figure illustrates the difference in contour-based loss between a collapsing tumour region and a volume-preserved tumour evaluated in image space and atlas space for a simplified (left) and a real world (right) example. The upper part of the figure shows the images with tumour contours (blue), the deformation vector fields (DVFs) in orange and the atlas. The lower part illustrates the calculated losses in image and atlas space for the DVF that collapses the tumour and the DVF that preserves the tumour volume. The plots demonstrate that an evaluation in atlas space favours the collapse of the tumour, while when evaluating the loss in image space, the volume preservation of the tumour is preferred.
• Figure 3: The expected (green) number of occurrences (assuming a uniform distribution) versus the measured (red) occurrences for all three centres and a combination of all three datasets. The stars indicate statistically significant differences, and the dashed line indicates the minimum number of samples required for the chi-square test. The figure shows counts only for brain areas with at least one metastasis in at least one of the three centres.
• Figure 4: Histogram of distance to the gray-white matter junction for melanoma metastasis barycentres and randomly sampled points. The y-axis represents the density of occurrences, while the x-axis shows the distance to the gray-white matter junction. The histogram includes two curves: one for metastasis barycentres (red line) and one for randomly sampled points (blue line). Compared to random sampling, metastasis barycentres exhibit a significantly higher density within a 1–2 mm distance from the gray-white matter junction, indicating a preferential spatial distribution of melanoma metastases in this region.
• Figure A.1: A simplified example illustrates that a collapse of the tumour region reduces the loss only when it is evaluated in atlas space, but not in image space. The orange arrows indicate the deformation vectors. When the tumour is preserved (left), no deformation is required. Due to a more irregular deformation field and therefore a higher regularization loss, the collapse of the tumour region results in a higher loss in image space compared to tumour preservation.