A Study on the Performance Comparison of Brain MRI Image-Based Abnormality Classification Models.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Neurology

Background:

• Class imbalance is a significant challenge in medical datasets, particularly for brain MRI.
• Limited availability of large-scale, real-world clinical data hinders the development of robust diagnostic models.
• Synthetic data generation offers a potential solution to overcome data scarcity issues.

Purpose of the Study:

• To develop and evaluate deep learning models for classifying normal versus abnormal brain MRI images.
• To compare the performance of deep learning models against traditional machine learning algorithms.
• To assess the impact of data augmentation and transfer learning on model performance.

Main Methods:

• A custom Convolutional Neural Network (CNN) and a ResNet-50 transfer learning model were fine-tuned.
• Experiments were conducted on a large synthetic dataset (10,000 images) after initial exploration on a small real-world dataset.
• Data preprocessing included normalization and augmentation; performance was compared with Support Vector Machines (SVM) and random forests.

Main Results:

• The ResNet-50 transfer learning model achieved approximately 95% accuracy and a high F1 score on the synthetic test set.
• The custom CNN model also demonstrated strong performance.
• Traditional methods (SVM, random forest) exhibited lower performance due to limitations in learning complex image features.

Conclusions:

• Deep learning, especially transfer learning with ResNet-50, is highly effective for brain abnormality detection in MRI.
• Synthetic data can be utilized for model training and evaluation when real-world data is scarce.
• Clinical applicability requires further validation using extensive real-world patient data.