[Paper Review] Biomechanical modelling of brain atrophy through deep learning

Mariana da Silva, Kara Garcia+4 more|arXiv (Cornell University)|Dec 14, 2020

Advanced Neuroimaging Techniques and Applications15 references4 citations

TL;DR

This paper proposes a differentiable, deep learning-based biomechanical model that simulates brain atrophy using prescribed atrophy maps as input, leveraging a Neo-Hookean material model and a U-Net architecture trained with a biomechanics-inspired loss function. The model achieves high accuracy in simulating longitudinal brain deformations, with atrophy prediction errors reduced to 5.1×10⁻⁵ (MSE) and residual deformations below 1.0×10⁻³ in image space, demonstrating rapid, biophysically plausible simulation for data augmentation and causal hypothesis testing.

ABSTRACT

We present a proof-of-concept, deep learning (DL) based, differentiable biomechanical model of realistic brain deformations. Using prescribed maps of local atrophy and growth as input, the network learns to deform images according to a Neo-Hookean model of tissue deformation. The tool is validated using longitudinal brain atrophy data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, and we demonstrate that the trained model is capable of rapidly simulating new brain deformations with minimal residuals. This method has the potential to be used in data augmentation or for the exploration of different causal hypotheses reflecting brain growth and atrophy.

Motivation & Objective

• To develop a differentiable, deep learning-based framework for simulating realistic brain atrophy patterns using prescribed local volume changes.
• To overcome the computational limitations of traditional finite element methods by replacing them with a learnable, end-to-end differentiable model.
• To enable rapid simulation of brain deformations with minimal residuals, suitable for data augmentation and causal modeling in neurodegenerative disease research.
• To integrate biomechanical plausibility into deep learning by embedding a Neo-Hookean strain energy function and anatomical constraints into the loss.
• To create a fully differentiable model that can be integrated into larger deep learning pipelines for explainable and causal medical imaging analysis.

Proposed method

• The model uses a U-Net to predict a displacement field from an input atrophy map, where each voxel's value represents relative volume change.
• The displacement field is applied via a Spatial Transformer layer to deform the original MRI image, enabling differentiable image transformation.
• The loss function incorporates a Neo-Hookean strain energy density model, with parameters for shear modulus (μ) and bulk modulus (K), and includes terms to enforce zero displacement outside the skull and fixed center of mass.
• The model is trained in an unsupervised manner using longitudinal T1-weighted MRI data from the ADNI dataset, with atrophy maps derived from FSL FNIRT registration Jacobians.
• The framework is implemented in PyTorch and supports both 2D and 3D inference, with the potential for extension to recurrent or generative architectures.
• Biomechanical constraints are enforced via a cost function that combines strain energy, background displacement penalty (λ₁), and center-of-mass fixation penalty (λ₂).

Experimental results

Research questions

• RQ1 Can a deep learning model accurately simulate brain atrophy patterns using only prescribed atrophy maps as input, while respecting biomechanical principles?
• RQ2 How does the performance of the proposed differentiable biomechanical model compare to traditional finite element methods in terms of accuracy and inference speed?
• RQ3 To what extent can the model generalize to new atrophy patterns not seen during training, particularly in complex or large-scale deformations?
• RQ4 Can the model be used to generate realistic synthetic data for data augmentation in neuroimaging studies?
• RQ5 How well does the model preserve tissue segmentation accuracy (e.g., CSF, WM, GM) under simulated atrophy?

Key findings

• The model achieved a mean squared error (MSE) of 5.1×10⁻⁵ in predicting atrophy maps from FNIRT-derived Jacobians, significantly lower than the baseline FNIRT method.
• The image-level MSE between simulated and ground-truth images was 1.0×10⁻³, with a standard deviation of 5.1×10⁻⁴, indicating high fidelity in image reconstruction.
• Dice scores for CSF, WM, GM, and DGM were 0.877, 0.921, 0.837, and 0.775 respectively, showing strong preservation of tissue segmentation across simulations.
• The model demonstrated robustness when trained on full-brain atrophy maps, with Dice scores for GM and WM remaining within 0.01 of the full-brain baseline, even when excluding the brain from the loss.
• The framework is fully differentiable and can be integrated into larger deep learning pipelines, enabling backpropagation through the biomechanical simulation process.
• The model successfully simulated brain atrophy with minimal residuals, suggesting strong generalization and physical plausibility of the learned deformation fields.