Classification of Noise-Free and Noise-Degraded Brain Tumour MRI Using Convolutional Neural Network

The early detection and treatment of brain tumours are essential, but diagnosing brain tumours accurately remains a significant challenge. This study presented a brain tumour classification model that utilized three convolutional neural network (CNN) models, namely VGG-16, SqueezeNet, and Inception-ResNet-V2, to classify brain tumour magnetic resonance images (MRI). The models were evaluated using established performance metrics such as precision, recall, F1 score, Matthew’s correlation coefficient (MCC), and accuracy. The models were executed using images from a Kaggle dataset, which included three types of brain tumours and one class of healthy brain images. The study identified the optimal hyperparameters, including a training time of 50 epochs, an SGD optimizer and a learning rate of 0.001, based on accuracy, loss, and other performance metrics. The VGG-16 and Inception-ResNet-V2 models achieved an accuracy of 99%, while SqueezeNet achieved 97%, alongside excellent precision, recall, F1 score, and MCC. The study also evaluated the impact of noise on the models’ performance, where image degradation in the form of self-introduced noise was included in addition to the original MRI dataset. The Inception-ResNet-V2 model was identified as the best-performing model due to its higher accuracy. The application of a filter to denoise the MRI image was also demonstrated in this work. In summary, the study’s findings highlighted the potential of AI-based systems to improve medical diagnostics.