Learning Spatio-Temporal Model of Disease Progression With NeuralODEs From Longitudinal Volumetric Data.

Area of Science:

• Medical image analysis
• Deep learning for disease progression modeling
• Computational anatomy

Background:

• Forecasting anatomical changes in age-related diseases is difficult.
• Accurate prediction can improve patient management and clinical trial design.
• Current methods struggle with modeling temporal pixel-level changes.

Purpose of the Study:

• To develop a deep learning method for predicting future anatomical changes from a single medical scan.
• To model the temporal evolution of age-related diseases.
• To evaluate the method on Geographic Atrophy and Alzheimer's Disease datasets.

Main Methods:

• Utilized Neural Ordinary Differential Equations (NeuralODEs) to model time-invariant physical processes.
• Incorporated domain-specific constraints and defined a temporal Dice loss.
• Trained and tested on retinal Optical Coherence Tomography (OCT) and brain Magnetic Resonance Imaging (MRI) datasets.

Main Results:

• Outperformed baseline models in predicting Geographic Atrophy growth.
• Achieved state-of-the-art results in predicting Alzheimer's Disease-induced brain ventricle changes on the TADPOLE dataset.
• Demonstrated effectiveness across different diseases and imaging modalities.

Conclusions:

• The proposed deep learning method effectively models and predicts temporal anatomical changes in age-related diseases.
• This approach holds promise for enhancing patient management and clinical trial efficiency.
• NeuralODEs provide a powerful framework for analyzing disease progression from medical imaging.