Lung nodule detection and classification based on geometric fit in parametric form and deep learning

This study presents an automated detection and classification method to facilitate the radiologists in the diagnosis process. The major problem in these systems is the inclusion of false positives in the results, which may lead to inaccurate diagnosis. A nodule detection and classification method is proposed that consists of four major phases. First, the lung region extraction is performed based on optimal gray level threshold that is computed by fractional-order Darwinian particle swarm optimization. Then, a novel nodule candidate detection method, based on geometric fit in parametric form incorporating the geometric properties of the nodules, is proposed. In the next phase, a hybrid geometric texture feature descriptor is created for better representation of the candidate nodules, which is a combination of 2D as well as 3D information about nodule candidates. Finally, a deep learning approach based on stacked autoencoder and softmax for feature reduction and classification is applied to reduce false positives. Performance analysis on the largest publically available dataset, Lung Image Database Consortium and Image Database Resource Initiative, depicts that the proposed method has significantly reduced the number of false positives to 2.8 per scan with a promising sensitivity of 95.6%. The results demonstrate the significance of the methodology in automatic lung nodule detection and classification. Furthermore, it will facilitate and provide assistance to radiologists in precise nodule detection.