Multi-Class Classification of Skin Tumour Using Convolution Neural Network (CNN) with Explainable Artificial Intelligence (XAI)

Sin, Yi Hong (2023) Multi-Class Classification of Skin Tumour Using Convolution Neural Network (CNN) with Explainable Artificial Intelligence (XAI). Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Full Text -Sin Yi Hong.pdf Restricted to Registered users only Download (11MB)

Abstract

Skin cancer is one of the most prevalent cancers in the world. The detection of skin tumours and classification of malignant and benign skin tumours are crucial to be performed as the chance of survival could be increased significantly when treatment is provided at the early stage. The diagnosis of skin tumours could be done visually but the accuracy of naked eye diagnosis for dermatologists is still low. The shortcoming motivates the introduction of artificial intelligence (AI) into skin tumour classification. The Convolution Neural Network (CNN) with the feature extraction characteristic is great for this purpose but it works as a “black box” which provides no justification for the decision. Hence. Explainable Artificial Intelligence (XAI) is necessary to introduce for explaining the decision made by the CNN model to ensure human understanding of the reason behind the decision. Four CNN models which are SqueezeNet, VGG16, ResNet50 and Inception ResNet-v2 were employed in this study and the hyperparameters were manipulated to optimise the performance of CNN models. The best-suit hyperparameter settings in this project were the 50 epochs, 0.1 dropout rate and the Adam optimizer with 0.001 learning rate. The pre-trained weight was applied and the dataset was normalised based on its mean and standard deviation. The best-performed CNN model for skin tumour classification is VGG16 with the high F1-score and Matthews Correlation Coefficient (MCC) which were 0.7125 and 0.6389 respectively. The performance of InceptionResnet-v2 was the worst among these four CNN models which were having 0.5705 F1 score and 0.4833 MCC score. The performance of ResNet50 and SqueezeNet was moderate which is not better than VGG16. The F1 score for these two models were 0.6737 and 0.7036 respectively and the MCC score of 0.6108 and 0.6282 respectively. Lastly, the LIME XAI model was implemented in this project. It was able to justify the prediction of the VGG16 model. The superpixels segmentation method of the LIME model was applied to the prediction images of the VGG16. The explanation of prediction of VGG16 was explained visually by the LIME model. The area of interest or called the features selected by the VGG16 model to justify and classify the image into a class was highlighted by the superpixels segmentation masking.