CLASSIFICATION OF BRAIN TUMOR USING HYBRID MODEL OF COSINE ANNEALING AND WEIGHTED SNAPSHOT ENSEMBLES

Abstract

The detection of brain tumors from MRI images is a critical initial step in brain cancer diagnosis, a task heavily dependent on the expertise of a radiologist. Consequently, there is a growing interest in developing automated diagnostic methods to assist radiologists. Hence, it minimizes the need for invasive procedures such as biopsies. Convolutional Neural Networks (CNNs) are recognized as a highly effective deep learning algorithm for accurate tumor identification and classification. While custom CNNs have proven effective in tumor classification, they often suffer from overfitting, hyperparameter sensitivity, and limited ensemble diversity, which limits their generalization performance. This study proposes a deep learning model based on a custom CNN that uses cosine annealing for learning-rate scheduling and a weighted snapshot ensemble of optimizers: Nadam, Adam, and Adamax. Cosine annealing is employed to mitigate fluctuations in validation performance during training, which can lead to overfitting, unstable training, misleading evaluation metrics, and increased risk of bias. The snapshot ensemble has been introduced to enhance the model’s classification performance. Each optimizer is trained independently, with two snapshots taken after 40 epochs to ensure proper convergence. The proposed model was trained and evaluated using the publicly available Figshare Dataset. Our approach achieved exceptional performance, with an accuracy, precision, recall, and F1-score of 97.9%, respectively. These results demonstrate the potential of our model to enhance automated brain tumor detection. Hence, supporting radiologists in any clinical setting.