Highly Efficient and Accurate Deep Learning-Based Classification of MRI Contrast on a CPU and GPU

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Classifying MRI images by contrast mechanism aids segmentation and automated processing.
• Inconsistent labeling schemes across institutions cause sorting ambiguity and manual errors.
• Automated classification is crucial for efficient archive management, retrieval, and training.

Purpose of the Study:

• To classify MRI images based on contrast mechanisms using transfer learning.
• To evaluate the performance of modified pretrained residual convolution neural networks.
• To assess the feasibility of CPU-based implementation for real-world deployment.

Main Methods:

• Transfer learning applied to pretrained residual convolution neural networks.
• Training and validation on 5169 MRI images across 10 classes, various vendors, and field strengths.
• Testing on a separate dataset of 2474 images; comparison of CPU vs. GPU performance.

Main Results:

• Achieved 99.76% accuracy in classifying MRI images by contrast mechanism.
• Training and validation completed in 36 minutes on a CPU.
• Heatmaps confirmed model's decision-making areas align with expert perception.

Conclusions:

• Transfer learning with residual networks offers highly accurate and efficient MRI image classification.
• CPU-based implementation is feasible for real-world applications, requiring minimal training time.
• The methodology is adaptable for other image classification tasks.