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

We developed a model that classifies normal and abnormal brain MRI images. This study initially referenced a small-scale real patient dataset (98 normal and 155 abnormal MRI images) provided by the National Institute of Aging (NIA) to illustrate the class imbalance challenge. However, all experiments and performance evaluations were conducted on a larger synthetic dataset (10,000 images; 5000 normal and 5000 abnormal) generated from the National Imaging System (NIS/AI Hub). Therefore, while the NIA dataset highlights the limitations of real-world data availability, the reported results are based exclusively on the synthetic dataset. In the preprocessing step, all MRI images were normalized to the same size, and data augmentation techniques such as rotation, translation, and flipping were applied to increase data diversity and reduce overfitting during training. Based on deep learning, we fine-tuned our own CNN model and a ResNet-50 transfer learning model using ImageNet pretrained weights. We also compared the performance of our model with traditional machine learning using SVM (RBF kernel) and random forest classifiers. Experimental results showed that the ResNet-50 transfer learning model achieved the best performance, achieving approximately 95% accuracy and a high F1 score on the test set, while our own CNN also performed well. In contrast, SVM and random forests showed relatively poor performance due to their inability to sufficiently learn the complex characteristics of the images. This study confirmed that deep learning techniques, including transfer learning, achieve excellent brain abnormality detection performance even with limited real-world medical data. These results highlight methodological potential but should be interpreted with caution, as further validation with real-world clinical MRI data is required before clinical applicability can be established.