Longitudinal Pooling & Consistency Regularization to Model Disease Progression From MRIs.

Area of Science:

• Neuroimaging
• Machine Learning
• Neurological Disorders

Background:

• Neurological diseases cause gradual brain deterioration, often studied using longitudinal MRI.
• Current deep learning models (CNNs and RNNs) for MRI analysis may yield clinically implausible classifications by neglecting disease progression.

Purpose of the Study:

• To develop a novel deep learning method for longitudinal MRI analysis that accounts for disease progression.
• To improve the accuracy and clinical plausibility of classifying brain changes over time in neurological conditions.

Main Methods:

• Proposed a new deep learning architecture combining Convolutional Neural Networks (CNNs) for feature extraction with a novel longitudinal pooling layer.
• Integrated a mechanism to enforce classification consistency across visits, aligning with disease progression.
• Evaluated the method on three large longitudinal neuroimaging datasets: ADNI, AUD, and NCANDA.

Main Results:

• The proposed method demonstrated superior performance compared to widely used longitudinal classification approaches across all three datasets.
• Achieved more accurate tracking of the impact of neurological conditions on brain structure over time.
• The novel longitudinal pooling layer effectively combined features across visits to model disease progression.

Conclusions:

• The developed method offers a more accurate and clinically relevant approach for analyzing longitudinal MRI data in neurological diseases.
• This technique advances the capability to track brain changes and disease progression using neuroimaging.
• The findings contribute to better understanding and monitoring of neurological conditions through advanced machine learning in neuroimaging.