Brain tumour detection and classification with hybrid deep learning models

Accurately classifying different forms of brain tumours is crucial for timely diagnosis and potentially lifesaving. High contrast greyscale brain images can be obtained non-invasively using magnetic resonance imaging (MRI), which is mostly utilised for tumour diagnosis. Significant improvements in medical imaging interpretation have resulted from the deep learning revolution brought about by the use of convolutional neural networks (CNN) in diagnostic systems. In order to categorise brain tumours into three sets—meningioma, glioma, and pituitary tumors—we use ResNets in conjunction with a transfer learning (TL) -based fine-tuning tool. Using the publicly available CE-MRI (Figshare) dataset, we fine-tune CNN five pre-trained models from the ResNets family, which includes ResNet18 through ResNet50. In our method, the pre-trained ResNet model is refined in two steps. The model is first initialised using weights taken from the ImageNet dataset. Then, in order to facilitate tumour classification, we incorporate more layers, such as top layers then a completely connected layer (FC). To evaluate the resilience of our fine-tuned ResNets against further pre-trained models, we run a number of tests. We also examine how data augmentation affects the test accuracy of the model. To comprehend the model’s decision-making process, we use Grad-CAM visualisation to look at the attention maps produced by the best model, which effectively highlights the locations of tumours in brain pictures. Our findings demonstrate that substantial performance gains are obtained when ResNet34 is used as the underlying framework. In particular, it was discovered that the complete test accuracy, recall, precision, and $\mathbf{F 1}$-score were $\mathbf{9 7. 6 1 \%}$, 97.45%, 98.74%, and 97.48%, respectively.