Classification of Colon Tumors based on Machine Learning Techniques and Deep Learning Techniques: A Comparative Study رسالة ماجستير

Abstract

The rapid evolution in technology nowadays especially in the medical sector has encouraged researchers to build Artificial Intelligence (AI) models to help in the diagnosis of critical diseases that threaten human lives such as cancers. Colon cancer is considered the third cause of death worldwide. The symptoms of colon cancer are not noticeable at the early stages but it could spread fast damaging human organs. The diagnosis of Colon cancer can be automated using Artificial Intelligence powerful models at early stages, with less time, more accuracy, and fewer costs. In this thesis, Machine Learning (ML) algorithms and Deep Learning (DL) algorithms were used to classify colon tumors using a local dataset from the Ministry of Health of symptoms and risk factors and another global colon cancer tissue images dataset. Eight different machine learning algorithms were used; Support Vector Machine (SVM), Naïve Bayes, Decision Tree (DT), K-nearest neighbor (KNN), Ensemble, Radial Basis Function Neural Network (RBFNN), Recurrent Neural Network (RNN), and Multi-Layer Perceptron Neural Network (MLPNN). Convolutional Neural Network (CNN), Visual Geometry Group -16 (VGG-16), and Visual Geometry Group 19 (VGG-19) were used for deep learning algorithms. A comparison of performance results for the eight machine learning algorithms was carried out to evaluate the model with the best accuracy for classifying colon cancer symptoms and risk factors. Another comparison VI between deep learning models is also done to evaluate the model with the best accuracy for classifying colon cancer tissue images. The results for the first experiment on a local dataset containing features selected by domain experts with machine learning algorithms showed that Ensemble and Naïve Bayes algorithms were able to achieve the best accuracy of 96.2%. Other algorithms; SVM, DT, KNN, RBF, RNN, and MLPNN achieved 87.80%, 92.4%, 79.2%, 66%, 92.5%, and 90.6% respectively. The same algorithms were applied o the local dataset after using mutual information feature selection. SVM, DC, Ensemble, and Recurrence Neural Networks achieved the best accuracy of 94.3%. For the second experiment using Deep learning techniques against the global dataset LC25000, VGG-19 was able to achieve the best classification accuracy of 98.94% while VGG-16 achieved 97.54% and CNN 83.94%. The ensemble algorithm is recommended for colon tissue tumor classification using the local dataset while VGG-19 is recommended with superior accuracy to classify colon tissue tumor images between cancerous and benign tumors.