Synergistic deep learning ensemble of DenseNet, ResNet, and Xception for accurate brain tumor diagnosis

Abstract

Brain tumor classification plays a pivotal role in medical diagnostics, treatment planning, and patient care. Currently, the biomedical domain has experienced a surge in demand of automated computer-aided diagnosis (CAD) systems for efficient diagnosis. Recently, deep learning (DL) architectures have experienced tremendous growth as prevailing tools for automated classification tasks. This study explores the synergistic potential of combining three state-of-the-art DL architectures i.e. densely connected convolutional network (DenseNet), residual network (ResNet), and extreme Inception (Xception) for brain tumor classification. Each architecture brings unique strengths, including feature reuse, skip connections, and depth-wise separable convolutions, respectively. This study proposes an ensemble methodology that capitalizes on the distinctive strengths of each architecture. DenseNet excels in feature reuse, ResNet employs skip connections to mitigate gradient issues, Inception captures multi-scale features, and Xception employs depth-wise separable convolutions for efficiency. By integrating predictions from these diverse models, our ensemble approach aims to achieve enhanced classification accuracy and robustness. After feature extraction, an improved feature selection method is introduced and finally, an ensemble classifier is introduced which uses majority voting method to achieve the classification task. The proposed approach has reported the average score as 0.99, 0.98, 0.99, 0.98, and 0.98 for precision, recall, specificity, sensitivity, and accuracy, respectively.