Classification of non-small cell lung cancer types using sparse deep neural network features

Most of the non-small cell lung cancer is clinically examined using CT/PET images. But an accurate diagnosis by the radiologist is difficult while classifying the type of non-small cell cancer, which may lead to misdiagnosis. Hence, a method is required to accurately identify different types of non-small cell lung cancers, such as adenocarcinoma and squamous cell carcinoma for providing proper treatment to patients. One of the practical and feasible solution is deep learning based method that has the ability to adapt and learn. However, most of the deep learning methods have complexity issues. Hence, some optimization is required to make the networks less complex. The objective of the work is to use less complex methods for classifying the non-small cell lung cancer. In this work, dense neural network (VGG-16 and Resnet-50) that has complex structures and sparse neural networks (inceptio v3) that are less complex are used. Deep learning methods are employed to obtain features from CT images and accurately classify non-small cell lung cancer. To evaluate the method, 60 adenocarcinoma patients and 60 squamous cell carcinoma patients are considered. The sensitivity, specificity, and accuracy of the Inception v3 network are found to be 96.66 %, 99.12 % and 98.29 % respectively. Observations indicate that the inception v3 model outperforms VGG-16 and ResNet-50. Also, the inception v3 network that is a sparse neural network has less computational overhead as compared to the other two networks. Sparse deep learning techniques may help radiologists accurately classify non-small cell lung cancer using CT images.