Detection and classification of pulmonary nodules using deep learning and swarm intelligence

Abstract

Cancer diagnosis is usually an arduous task in medicine, especially when it comes to pulmonary cancer, which is one of the most deadly and hard to treat types of that disease. Early detecting pulmonary cancerous nodules drastically increases surviving chances but also makes it an even harder problem to solve, as it mostly depends on a visual inspection of tomography scans. In order to help improving cancer detection and surviving rates, engineers and scientists have been developing computer-aided diagnosis systems, similar to the one presented in this paper. These systems are used as second opinions, to help health professionals during the diagnosis of numerous diseases. This work uses computational intelligence techniques to propose a new approach towards solving the problem of detecting pulmonary carcinogenic nodules in computed tomography scans. The applied technology consists of using Deep Learning and Swarm Intelligence to develop different nodule detection and classification models. We exploit seven different swarm intelligence algorithms and convolutional neural networks, prepared for biomedical image segmentation, to find and classify cancerous pulmonary nodules in the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) databases. The aim of this work is to use swarm intelligence to train convolutional neural networks and verify whether this approach brings more efficiency than the classic training algorithms, such as back-propagation and gradient descent methods. As main contribution, this work confirms the superiority of swarm-trained models over the back-propagation-based model for this application, as three out of the seven algorithms are proved to be superior regarding all four performance metrics, which are accuracy, precision, sensitivity, and specificity, as well as training time, where the best swarm-trained model operates 25% faster than the back-propagation model. The performed experiments show that the developed models can achieve up to 93.71% accuracy, 93.53% precision, 92.96% sensitivity, and 98.52% specificity.