From Raw Pixels to Recurrence Image for Deep Learning of Benign and Malignant Mediastinal Lymph Nodes on Computed Tomography

Lung cancer causes the most cancer deaths worldwide and has one of the lowest five-year survival rates of all cancer types. It is reported that more than half of patients with lung cancer die within one year of being diagnosed. Because mediastinal lymph node status is the most important factor for the treatment and prognosis of lung cancer, the aim of this study is to improve the predictive value in assessing the computed tomography (CT) of mediastinal lymph-node malignancy in patients with primary lung cancer. This paper introduces a new method for creating pseudo-labeled images of CT regions of mediastinal lymph nodes by using the concept of recurrence analysis in nonlinear dynamics for the transfer learning. Pseudo-labeled images of original CT images are used as input into deep-learning models. Three popular pretrained convolutional neural networks (AlexNet, SqueezeNet, and DenseNet-201) were used for the implementation of the proposed concept for the classification of benign and malignant mediastinal lymph nodes using a public CT database. In comparison with the use of the original CT data, the results show the high performance of the transformed images for the task of classification. Three pretrained convolutional neural networks that are AlexNet, SqueezeNet, and DenseNet201 were trained and tested with the transformed images. Classification accuracies and areas under the receiver operating characteristic curve obtained from the ten-fold cross-validation are 93% and 0.97, 96% and 0.99, and 100% and 1 for the SqueezeNet, AlexNet, and DenseNet201, respectively. The proposed method has the potential for differentiating benign from malignant mediastinal lymph nodes on CT, and may provide a new way for studying lung cancer using radiology imaging.