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

Nanna Wielenberg, Ilinca Popp+7 more|arXiv (Cornell University)|Feb 13, 2026

Cutaneous Melanoma Detection and Management0 citations

TL;DR

Introduces a fully differentiable, preprocessing-free deep learning atlas-registration framework that aligns MBM to a common atlas while preserving metastases, enabling robust cohort analyses across multi-centre data.

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.

Motivation & Objective

• Reduce inter-patient anatomical variability to enable reproducible cohort-level analyses of melanoma brain metastases (MBM).
• Register pathological brain images to a common atlas without preprocessing or lesion masking.
• Preserve metastatic tissue during registration and handle missing correspondences robustly.
• Validate the framework on multi-centre MBM cohorts and analyze spatial distribution patterns.

Proposed method

• Use a fully differentiable deformable registration framework that maps each subject to an atlas via a deformation field computed by a U-net that outputs a stationary velocity field.
• Employ a sampling module that applies transformations through a learnable sampling grid to avoid image degradation from resampling.
• Train a general model on Learn2Reg healthy datasets, then perform one-shot over-fitting (Learn2Reg-OF and Clin1-Clin3-OF) to individual cases for higher accuracy.
• Use two loss terms: a label-map based similarity loss (via distance-transformed labels and NCC) and a regularization/volume-preservation term, plus an optional volumetric loss for handling missing correspondences during over-fitting.
• Evaluate registration with Dice, Hausdorff distance, and ASSD, plus Jacobian-based plausibility and tumour-volume changes.
• Publicly available implementation of the framework with atlas-based MBM analysis across three centres.

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 ( $T_{i}$ , $T_{j}$ )

Experimental results

Research questions

• RQ1 Can a deep-learning deformable registration framework align MBM to a common atlas without preprocessing or lesion masks while preserving tumor tissue?
• RQ2 Does the method enable robust, reproducible cohort-level spatial analyses across multi-centre MBM datasets?
• RQ3 What are the spatial distribution patterns of MBM when mapped to an atlas, and how do they relate to known neuroanatomical and perfusion landmarks?
• RQ4 How does one-shot over-fitting to individual cases compare to a general model in terms of accuracy and preservation of pathological tissue?

Key findings

• Registration accuracy remained high across datasets (DSC 0.89–0.92, HD 6.79–7.60 mm, ASSD 0.63–0.77 mm) with preserved metastatic volumes.
• MBM showed over-representation in cerebral cortex and putamen and under-representation in white matter after atlas mapping, with metastases preferentially near the gray–white matter junction.
• Arterial territories did not show a significantly increased MBM frequency after correcting for regional volume.
• Volume-preserving loss improved tumour preservation during over-fitting (e.g., Clin1-OF DSC 0.89, HD 7.27 mm, ASSD 0.77 mm; tumour volume factor ~0.82, Jacobian ratio ~0.91).
• The framework achieved strong registration while avoiding preprocessing and lesion masking, enabling reproducible cohort analyses across centres.

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 contou