Revisiting Classical and Deep Learning Models for Multi-Class Brain MRI Classification under Severe Data Constraints

Abstract

Context: Deep learning has great potential in neuroimaging but its efficacy is often constrained by the availability of large, standardized datasets. Data scarcity and heterogeneity are still major issues in many biomedical research settings. Objective: In this work, we analyze the performance of traditional machine learning models and deep neural networks for five-class brain MRI classification (Alzheimer’s illness, ischemic stroke, meningitis, brain tumor, and normal brain) using a small heterogeneous dataset of 2D slices. Method: We tested Local Binary Patterns (LBP) with K-Nearest Neighbors (KNN) and Gray-Level Co-occurrence Matrix (GLCM) with KNN, Support Vector Machine (SVM), tuned multilayer perceptron (MLP), lightweight convolutional neural network (CNN) and transfer learning model based on ResNet-50. The study focuses on model behavior, generalization, and limitations in constrained data scenarios, rather than just correctness. Results: The studies indicate that CNNs exploit spatial characteristics but generalize poorly, whereas SVM is quite accurate due to dataset bias. ResNet-50 demonstrates early saturation, indicating the disparity between model size and dataset scale. Conclusions: This work highlights the need for careful interpretation of numerical accuracy in tiny medical imaging datasets, and provides practical insights into how different model families respond to real-world data restrictions. These results provide insights for biomedical imaging studies, informing model choice in resource-limited settings, and providing opportunities for future clinical applications in settings with limited data availability.