Brain MRI analysis using a deep learning based evolutionary approach.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• Convolutional neural network (CNN) models show promise in medical image analysis, but their decision-making processes remain unclear.
• Understanding the internal workings of deep learning models is crucial for reliable medical applications.

Purpose of the Study:

• To develop a visualization technique for interpreting three-dimensional convolutional neural networks (3D-CNNs) in brain MRI analysis.
• To identify brain regions critical for classification tasks using a novel genetic algorithm-based brain masking (GABM) method.

Main Methods:

• A 3D-CNN was trained for brain MRI classification.
• A genetic algorithm (GA) with a new chromosome encoding approach was used to discover 'knowledgeable' brain regions utilized by the 3D-CNN.
• The framework was evaluated on Alzheimer's Disease Neuroimaging Initiative (ADNI) and Autism Brain Imaging Data Exchange (ABIDE) datasets.

Main Results:

• The 3D-CNN achieved 5-fold classification accuracies of 0.85 on ADNI and 0.70 on ABIDE.
• The GABM method identified 6-65 knowledgeable regions in ADNI and 15-75 in ABIDE.
• The visualization technique provided insights into feature extraction and improved model performance in some cases.

Conclusions:

• The proposed GABM method enhances the interpretability of 3D-CNNs in brain MRI analysis.
• This approach can help identify critical brain regions for disease classification.
• The method offers potential for improving both model understanding and classification accuracy.