EfficientNet and multi-path convolution with multi-head attention network for brain tumor grade classification

Grade classification is a challenging task in brain tumor image classification. Contemporary models employ transfer learning technique to attain better performance. The existing models ignored the semantic features of a tumor during classification decisions. Moreover, contemporary research requires an optimized model to exhibit better performance on larger datasets. Thus, we propose an EfficientNet and multi-path convolution with a multi-head attention network for the grade classification. We used a pre-trained EfficientNetB4 in the feature extraction phase. Then, a multi-path convolution with multi-head attention network performs a feature enhancement task. Finally, features obtained from the above step are classified using a fully connected double dense network. We utilize TCIA repository datasets to generate a three-class (normal/low-grade/high-grade) classification dataset. Our model achieves 98.35% accuracy and 97.32% Jaccard coefficient. The proposed model achieves superior performance than its competing models in all key metrics. Further, we achieve similar performance on a noisy dataset.