A new Multi-scale Dilated deep ResNet model for Classification of Lung Nodules in CT images

Lung nodules are the main indication of lung cancer. The main challenge for the radiologist is to diagnose and identify malignancies in computation tomography images. Different deep learning strategies for the treatment of lung cancer boost the efficiency of computer-aided diagnosis systems continually. Conventional neural networks use the focus layer to reduce resolution gradually but lack the ability to capture the features of small but critical pulmonary nodules. To handle this problem, we propose a new multi-scale dilated deep ResNet (MsDdR) model that helps classify lung nodules between benign and malignant. In this model, we apply different dilation rates, i.e. (3,5,7 and 9), on the input CT images to gain more relative information of the nodules. Then the images are transferred to the deep ResNet. Furthermore, deep ResNet is developed by merging residual learning with migrating learning by ignoring conventional image processing and by carrying over a 50 layers ResNet structure. The performance of the proposed model is analysed using several assessment matrices of accuracy, specificity, and AUC. The public accessible lung image database consortium dataset is used to evaluate the performance in the experiment. Experimental results prove the performance of the model with an accuracy of 91.26% and AUC 0.957.