Assessing the Performance of Deep Learning Models for Colon Polyp Classification using Computed Tomography Scans
Khadija Hicham, Sara Laghmati, Soufiane Hamida, Asmae El Ghazi, Amal Tmiri, Bouchaib Cherradi
Abstract
the diagnosis of Colorectal and Rectum Cancer (CRC) is a global concern as it is the third most commonly diagnosed cancer. Early detection and treatment of polyps can prevent the development of colon cancer. To assist with this, Computed Tomography (CT) scans are used to produce three-dimensional images of the interior of the colon. Deep Learning techniques, such as Convolutional Neural Networks (CNNs), have the potential to provide valuable support to radiologists in the early detection of colon polyps. In this study, we developed and trained three Deep Learning models, VGG16, VGG19, and 3DCNN15, from scratch to classify CT scans based on the presence or absence of colon polyps. The dataset used for this study was a CT colonography dataset, which consisted of 3D images of the colon's interior. The proposed system was evaluated using a confusion matrix and four evaluation metrics: accuracy, precision, recall, and F1 score. Our findings suggest that while the models can assist radiologists in classifying polyps in 3D scans with an accuracy of 76.7%, there is room for improvement.