Deep Convolutional Neural Network-based Automatic Detection of Brain Tumour

Deep convolutional neural networks (DCNNs) have been extensively studied for different types of detection and classification in the field of biomedical image processing. Many of them have produced results that are on par with or even better than those of radiologists and neurologists. But, the challenge to get good results from such DCNNs is the requirement of large dataset. In this paper, a unique single-model based approach for classifying brain tumours on small dataset is presented in this study. A modified DCNN called the RegNetY-3.2G is used, integrated with regularization DropOut and DropBlock to prevent over-fitting. Furthermore, an improved augmentation technique called the RandAugment is used to lessen the problem of small dataset. Lastly, MWNL (Multi-Weighted New Loss) method and end to end CLS (cumulative learning strategy) is used to address the problem of unequal size of sample, complexity in the classification and to lessen the effect of aberrant samples on training.