Smart Data Augmentation for Multiclass Brain Tumor Detection with Imbalanced Data

Dealing with class imbalance is a significant issue in classification tasks that often leads to lower prediction performance. Many data augmentation methods have been suggested to tackle this problem, but their effectiveness can vary depending on the dataset. Utilizing augmentation methods can improve the learning and generalization abilities of the proposed convolutional neural network (CNN) model. However, it is challenging to find a single data augmentation method that can effectively address class imbalance problems across all scenarios. In this paper, we present a technique known as smart data augmentation (SDA) that is aimed at effectively improving classification performance by enhancing imbalanced data. SDA involves an equation that can be used to create an augmentation technique, combining different sampling methods for addressing multiclass imbalance and offering easy customization options. We perform the data balancing strategy including undersampling, oversampling, and SDA-based sampling. We evaluate our proposed framework using a CNN model to improve the detection of brain tumors in imbalanced data. Our experiments demonstrate that SDA-based sampling improves brain tumor detection performance by 3.546% in accuracy, 3.310% in precision, 3.432% in recall, and 3.371 % in Fl-score compared to the imbalanced data scenario. The findings suggest that utilizing SDA together with CNN can improve the classification performance of brain tumor detection with imbalanced data.