Automatic spatial estimation of white matter hyperintensities evolution in brain MRI using disease evolution...

Area of Science:

• Neuroimaging
• Medical Artificial Intelligence
• Radiology

Background:

• White matter hyperintensities (WMH) are key indicators of small vessel disease and can change over time.
• Predicting WMH evolution is complex due to unknown clinical risk factors, making it a non-deterministic task.

Purpose of the Study:

• To develop and evaluate a deep learning model, the Disease Evolution Predictor (DEP), for predicting WMH evolution from baseline to 1-year follow-up.
• To explore both supervised (DEP-UResNet) and unsupervised (DEP-GAN) deep learning approaches for WMH evolution prediction.
• To investigate the impact of auxiliary inputs, such as Gaussian noise, on the model's ability to handle the non-deterministic nature of WMH changes.

Main Methods:

• Proposed two DEP models: DEP-UResNet (supervised) and DEP-GAN (unsupervised).
• Incorporated auxiliary inputs, including Gaussian noise, baseline WMH, and stroke lesion loads, to simulate non-deterministic factors.
• Evaluated model performance using clinical assessments and ablation studies.

Main Results:

• Fully supervised DEP-UResNet generally outperformed unsupervised DEP-GAN.
• A semi-supervised DEP-GAN achieved performance comparable to DEP-UResNet and excelled in clinical evaluation.
• Auxiliary inputs, particularly Gaussian noise, significantly improved DEP model performance across architectures.

Conclusions:

• Deep learning models, especially with semi-supervised approaches and auxiliary inputs like Gaussian noise, can effectively predict WMH evolution.
• This study represents a novel approach to modeling the non-deterministic nature of WMH changes using AI.