A Novel Hybrid Approach for Brain Tumour MRI Segmentation Leveraging Transfer Learning

Integrating multiple deep learning architectures often significantly improves the performance of a model. This approach is particularly beneficial for challenging tasks like brain tumour detection. This paper provides a detailed examination of a novel DenseNet-ResNet fusion model that leverages DenseNet's strength in feature extraction and ResNet's robust learning process. This helps in overcoming the limitations of each separate model and, hence, leads to increased accuracy and improved training efficiency. Additionally, various augmentation techniques were applied to the data to improve the performance; these included brightness adjustments, rotation, shifting, shear transformation, zooming, and horizontal flipping. We also employed histogram equalization, to improve contrast, making the image clearer for visualizing the tumours. Furthermore, image registration algorithms were embedded for the alignment of flipped images, and Grad-CAM visualization was utilized for a more comprehensive evaluation of the results. Our results have shown that the DenseNet-ResNet fusion model not only gives higher classification accuracy but is also better at generalization.