Automatic Lung Cancer Detection and Diagnosis Using Hand Crafted and Deep Learning Features

This paper presents a lung nodule detection and classification system which utilizes a combination of hand crafted and deep learning features. Hand crafted features were obtained from modified methods of bag of frequencies, and taxonomic indices. We included a robust radius estimation algorithm that resulted in an average error of 1.29 pixels. Hand crafted features were obtained from 3D low dose CT scans pertaining to lung cancer patients as well as control subjects, and we used the following specific features: the suspected nodule location, radius of the nodule, spectral signatures, taxonomic diversity and distinctness. Deep learning features were obtained with Inception-v3, a pretrained network trained on ImageNet, that is currently the state of the art Convolutional Neural Network (CNN) architecture for the ILSVRC challenge. Our work aims to expand upon existing detection and classification methods to improve lung cancer screening accuracies. Promising results indicated the superiority of combined features for lung tumor detection and diagnosis. Keywords—lung nodules, computed tomography (CT), computeraided detection, nodule detection, nodule characterization, pulmonary nodules, deep learning