Lung nodule classification using deep feature fusion in chest radiography.

Lung nodules are small, round, or oval-shaped masses of tissue in the lung region. Early diagnosis and treatment of lung nodules can significantly improve the quality of patients' lives. Because of their small size and the interlaced nature of chest anatomy, detection of lung nodules using different medical imaging techniques becomes challenging. Recently, several methods for computer aided diagnosis (CAD) were proposed to improve the detection of lung nodules with good performances. However, the current methods are unable to achieve high sensitivity and high specificity. In this paper, we propose using deep feature fusion from the non-medical training and hand-crafted features to reduce the false positive results. Based on our experimentation of the public dataset, our results show that, the deep fusion feature can achieve promising results in terms of sensitivity and specificity (69.3% and 96.2%) at 1.19 false positive per image, which is better than the single hand-crafted features (62% and 95.4%) at 1.45 false positive per image. As it stands, fusion features that were used to classify our candidate nodules have resulted in a more promising outcome as compared to the single features from deep learning features and the hand-crafted features. This will improve the current CAD method based on the use of deep feature fusion to more effectively diagnose the presence of lung nodules.