VoxelHop: Successive Subspace Learning for ALS Disease Classification Using Structural MRI.

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Neurological disease classification

Background:

• Deep learning models for medical imaging face challenges with small datasets and lack of transparency.
• Existing convolutional neural network (CNN) architectures often require large datasets and significant memory.

Purpose of the Study:

• To introduce VoxelHop, a successive subspace learning model for accurate classification of Amyotrophic Lateral Sclerosis (ALS).
• To address the limitations of deep learning models in terms of dataset size and interpretability in clinical settings.

Main Methods:

• VoxelHop utilizes sequential neighborhood expansion, unsupervised and label-assisted subspace dimension reduction, and feature concatenation.
• The model employs a modular and transparent structure without backpropagation, suitable for 3D imaging data.
• Classification is performed by comparing features between control and patient groups.

Main Results:

• VoxelHop achieved 93.48% accuracy and an AUC score of 0.9394 in differentiating ALS patients from controls.
• The model demonstrated robustness and effectiveness with a small dataset (20 controls, 26 patients).
• Evaluations showed superiority over state-of-the-art 3D CNN classification approaches.

Conclusions:

• VoxelHop offers an effective and robust solution for ALS classification using T2-weighted structural MRI.
• The model's transparency and efficiency make it suitable for clinical adoption, especially with limited data.
• The VoxelHop framework is generalizable to other classification tasks across different imaging modalities.