Abstract
White Matter Hyperintensities (WMH) are key indicators of small vessel disease, stroke, and dementia. Predicting their evolution—whether they grow, stabilize, or regress—is vital for clinical intervention but remains difficult due to high variability. We present the Disease Evolution Predictor GAN (DEP-GAN), which leverages unsupervised Irregularity Maps (IM) and a dual-critic architecture to simulate voxel-level changes. By incorporating Feature-wise Linear Modulation (FiLM) to model inherent uncertainty, DEP-GAN significantly outperforms traditional models in volumetric accuracy and growth detection.
Introduction
White Matter Hyperintensities (WMH) are bright spots visible on brain MRIs that serve as critical biomarkers for underlying conditions such as stroke and dementia. A significant challenge in clinical neurology is that these lesions do not follow a linear path; they can grow, remain stable, or even regress over time. Predicting this evolution at a voxel level is notoriously difficult due to the poorly understood biological drivers and high inter-patient variability.
Traditional predictive models often focus on general disease progression rather than specific structural changes. To address this, researchers at the University of Edinburgh developed the Disease Evolution Predictor GAN (DEP-GAN), a deep learning model designed to estimate the future size and location of WMH, providing a crucial tool for designing therapeutic interventions.
Methods
The DEP-GAN framework introduces several innovations to handle the complexity of brain lesion evolution:
Irregularity and Evolution Mapping
Instead of relying on biased manual labels, the team utilized Irregularity Maps (IM) generated via an unsupervised method (LOTS-IM). These maps capture subtle textures and "irregularities" compared to normal tissue. By subtracting a baseline IM from a follow-up IM, they created a Disease Evolution Map (DEM), where positive values indicate progression and negative values indicate regression.
DEP-GAN Architecture
The model employs a Dual-Critic Architecture. While standard GANs use one discriminator, DEP-GAN uses two:
- Critic D: Ensures the predicted follow-up image is anatomically realistic.
- Critic C: Ensures the process of change (the DEM) follows biological rules.
To account for the non-deterministic nature of the disease, the generator is modulated with random noise using Feature-wise Linear Modulation (FiLM). This allows the model to represent a spectrum of possible outcomes rather than a single "average" prediction.
Results and Discussion
The study utilized MRI data from 152 stroke patients. DEP-GAN-2C (the dual-critic version) significantly outperformed baseline models like WGAN-GP and VA-GAN across several metrics:
- Volumetric Accuracy: The model estimated follow-up WMH volume with a mean error of -1.91 ml, a vast improvement over the -16.89 ml error seen in standard WGAN-GP.
- Evolution Prediction: It achieved 72.01% accuracy in predicting the direction of change (growth vs. regression).
- Growth Detection: The model was particularly robust in predicting lesion growth, reaching 89.47% accuracy.
- Spatial Precision: Higher Dice Similarity Coefficients (DSC) indicated superior accuracy in predicting the exact location of future lesions.
The success of the dual-critic approach suggests that constraining the process of change is as important as the final output. Furthermore, the use of unsupervised IMs makes the model highly scalable, as it does not require labor-intensive manual segmentation.
Conclusion
DEP-GAN represents a significant step forward in modeling the complex dynamics of small vessel disease. While currently limited by a small 2D dataset, the framework demonstrates the potential of GANs to handle biological uncertainty. Future developments will focus on incorporating clinical biomarkers (e.g., blood pressure, genetics) and expanding the architecture to 3D to better capture the volumetric nature of brain lesions.
Summary updated