Three-Dimensional Deep Learning with Routine Brain Magnetic Resonance Imaging and Clinical Data for Identification of...

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Secondary progressive multiple sclerosis (SPMS) is a transition from relapsing-remitting multiple sclerosis (RRMS).
• Distinguishing between SPMS and RRMS is crucial for timely diagnosis and management.
• Current understanding of individual differences between these MS phenotypes is limited.

Purpose of the Study:

• To investigate the efficacy of 3D deep learning frameworks (VGG19, ResNet152, DenseNet-121) in differentiating SPMS from RRMS.
• To identify significant brain regions contributing to this differentiation using model explanation techniques.
• To assess the impact of incorporating clinical variables alongside MRI data.

Main Methods:

• Utilized a dataset of 140 participants (70 RRMS, 70 SPMS) with routine clinical and brain MRI data.
• Employed 3D deep learning models (VGG19, ResNet152, DenseNet-121) trained on T1-weighted, T2-weighted, and FLAIR MRI sequences.
• Applied 3D Grad-CAM for model explanation and ablation studies to pinpoint significant brain areas.
• Incorporated optional clinical variables (n=6) with MRI data for enhanced model performance.

Main Results:

• VGG19 emerged as the best-performing 3D framework for differentiating SPMS from RRMS.
• Models combining MRI and clinical data showed equivalent or improved performance compared to MRI-only models (AUC up to 0.92).
• Significant brain regions identified include bilateral frontal lobes, left occipital and temporal lobes, and cerebellum.

Conclusions:

• 3D deep learning models, particularly VGG19, show significant potential for distinguishing SPMS from RRMS using routine data.
• The integration of 3D Grad-CAM can help identify key brain regions, potentially uncovering new biomarkers for disease worsening.
• This approach could enhance the diagnosis and management of multiple sclerosis phenotypes.