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

Yong, Weng Seen (2023) Classification of Non-Degraded and Noise-Degraded Brain Tumour MRI Using Convolutional Neural Network. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Full Text -Yong Weng Seen.pdf Restricted to Registered users only Download (7MB)

Abstract

The brain tumour is one of the most prevalent tumours in the human body. Therefore, intervention is necessary during the early stage to detect and cure brain tumours. The diagnosis of brain tumours by professionals is subjected to the risk of misdetection due to human errors induced by heavy workload and exhaustion. The determination of tumour type also required a high level of medical expertise from professionals. Such obstacles drive the need for a brain tumour classification system, aided by artificial intelligence (AI) to perform automatic brain tumour classification. Hence, upon reviewing various Convolutional Neural Network (CNN) architectures and reported literature on their applications in biomedical imaging, three CNN models for brain tumour classification namely VGG-16, SqueezeNet, and Inception-ResNet-V2 are proposed to be implemented in this study. The hyperparameters would be manipulated to optimise the performance of the CNN models. The performance of the different architectures with the selected hyperparameters would be evaluated and compared according to the well-established evaluation metrics for deep learning (DL) namely precision, recall, F1 score, Matthew’s correlation coefficient (MCC) and accuracy. The system will be created using Python programming language and executed in Kaggle using images obtained from Brain Tumour Magnetic Resonance Image (MRI) dataset which consists of 4 classes: 3 types of brain tumour and 1 class of healthy brain. In addition, image degradation in terms of self-introduced noise would be included in the MRI dataset to observe the effects of image degradation in brain tumour MRI classification. Based on accuracy, loss, and other performance matrix, the selected hyperparameters are, 0 dropout layer, 50 number of epochs, SGD optimizer with 0.95 momentum rate and 0.001 learning rate. The highest accuracy in brain tumour classification was 99% achieved by VGG-16 and Inception-ResNet-V2, whereas SqueezeNet achieved an accuracy of 98%. These CNN architectures have also presented excellent precision, recall, F1 score and MCC. After considering the performance metrics, the VGG-16 model is proposed as the final model to be implemented in the brain tumour classification system. Gaussian noise, salt-and-pepper noise were introduced into the augmented dataset to determine the effect of noise to the classification performance of the CNN models. After introducing noise, Inception-ResNet-V2 was selected as the best performed model due to its higher accuracy than the other 2 CNN models. However, after noise filter was added, the noise filters successfully mitigated the impact of noise, allowing VGG-16 to excel in accurately classifying the images.