Lung cancer detection using multi-layer neural networks with independent component analysis: A comparative study of training algorithms

The present paper presents a computer-aided design (cad) system that detects lung cancer. lung cancer detection uses multi-layer (ml),neural networks (nns) and independent component analysis (ica). ica aims to speed the detection by decreasing the number of features. the ml nns classifier is trained by gradient descent algorithm (traingd),gradient descent with momentum (traingdm),gradient descent with variable learning rate and momentum (traingdx),resilient back propagation (trainrp),fletcher-reeves update (traincgf),polak and ribiere (traincgp),powell and beale restarts (traincgb),scaled conjugate gradient algorithm (trainscg),quasi newton bfgs (trainbfg),one step secant algorithm (trainoss) and levenberg-marquardt (trainlm). the detection algorithm is tuned to determine the existence of cancer in real computerized tomography (ct) images and it is validated,trained & tested using 460 ct images,350 of them belong to lung cancer patients in jordanian hospitals. the presence of cancer in these images is labeled by experts. the present paper investigates the performance of the ml nn classifier trained by these training algorithms with ica feature extraction. results reveal the robustness of the detection algorithm for real ct images. among the 11 training algorithms,levenberg-marquardt achieves a classification accuracy of 100% with least number of ica features. © 2017 jordan journal of biological sciences.