Comparison of deep learning convolutional neural network (CNN) architectures for CT lung cancer classification

Lung cancer has become one of the most common deaths amongst the cancer patients. World Health Organisation states that lung cancer is the second most fatal cancer all over the world in 2014. Alarmingly, most of the lung cancer patients are diagnosed at the later stages where the cancer has spreads. Thus, early screening via Computed Tomography scan particularly among active smokers is encouraged. Manual diagnosis of the cancer is made feasible through the integration of Computer Aided Diagnosis system. For the past few years, deep learning method leads most of the artificial based intelligence applications including CAD systems. This paper aims to investigate the performance of five newly established Convolutional Neural Network architectures; GoogleNet, SqueezeNet, DenseNet, ShuffleNet and MobileNetV2 to classify lung tumours into malignant and benign categories using LIDC-IDRI datasets. Their performances are measured in terms of accuracy, sensitivity, specificity and area under the curve of the receiver operating characteristic curve. Experimental results show that GoogleNet is the best CNN architecture for CT lung tumour classification wih an accuracy of 94.53%, specificity 99.06%, sensitivity of 65.67% and AUC 86.84%.