Comparative Analysis of Data Augmentation Techniques for Enhancing Generalizability in Brain Tumor Classification with Limited Datasets

Deep learning models for brain tumor classification often face challenges related to generalizability, primarily due to the limited availability of labeled medical data. While data augmentation techniques offer a promising solution by expanding datasets artificially and potentially improving model performance, existing research lacks a comprehensive comparative analysis of their effectiveness in enhancing model generalizability. Furthermore, the literature lacks extensive discussions on the performance of models trained on limited datasets and their ability to generalize to unseen, potentially larger datasets. To address these gaps, this study investigates the generalizability of various data augmentation techniques for brain tumor classification with limited data. Specifically, we focus on two key problems: (1) the comparative analysis of model generalizability to unseen datasets, and (2) the comparison of augmentation techniques and their impact on model generalizability. We train a ResNet50 model on a small dataset and evaluate its performance on multiple datasets, augmented using techniques including geometric transformations, SMOTE, ADASYN, and GAN-based augmentation. Through an analysis of classification performance metrics such as accuracy, precision, recall, and F1-score, we aim to identify the most effective techniques for enhancing the generalizability of models trained on limited datasets.