Training and testing a CNN-based engine for brain MRI scan classification and segmentation

Biomedical imaging data (from X-rays, CT scans, MRIs) is a rich and crucial source of information for diagnosing and treating patients, yet the large volume of visual data generated can be challenging even for the most experienced clinical professionals to handle and use for diagnosis with high accuracy. With the traditional diagnosis approaches, patients usually face high costs too. Sometimes, abnormalities remain unspotted, causing delays in intervention which can be the difference between saving and losing patient lives. Most traditional radiomics studies use hand-crafted feature extraction techniques, such as texture analysis, followed by conventional machine learning classifiers, such as random forests and support vector machines (SVMs). In this paper, we establish that X-ray, MRI or CT-image classification using convolutional neural networks (CNNs) could be an efficient, cost-effective and fast approach for diagnosis and interpretation. The research is focused on training a CNN algorithm to develop diagnostic analysis power using 250 brain MRI images and then testing the accuracy and predictive power of the developed CNN algorithm on 5 images. We have shown the superiority of CNN models in image classification compared to traditional ML Classifiers that used an extensive ablation study. We have then compared a selected set of existing models with the new model we have built based on the EfficientNet v2-S architecture with a high classification accuracy of 98% – showing a high predictive power to provide various differential diagnoses from the brain MRI scans.