High Throughput Lung and Lobar Segmentation by 2D and 3D CNN on Chest CT with Diffuse Lung Disease

Abstract

Deep learning methods have been widely and successfully applied to the medical imaging field. Specifically, fully convolutional neural networks have become the state-of-the-art supervised segmentation method in a variety of biomedical segmentation problems. Two fully convolutional networks were proposed to sequentially achieve accurate lobar segmentation. Firstly, a 2D ResNet-101 based network is proposed for lung segmentation and 575 chest CT scans from multicenter clinical trials were used with radiologist approved lung segmentation. Secondly, a 3D DenseNet based network is applied to segment the 5 lobes and a total of 1280 different CT scans were used with radiologist approved lobar segmentation as ground truth. The dataset includes various pathological lung diseases and stratified sampling was used to form training and test sets following a ratio of 4:1 to ensure a balanced number and type of abnormality present. A 3D CNN segmentation model was also built for lung segmentation to investigate the feasibility using current hardware. Using 5-fold cross validation a mean Dice coefficient of 0.988 ± 0.012 and Average Surface Distance of 0.562 ± 0.49 mm was achieved by the proposed 2D CNN on lung segmentation. 3D DenseNet on lobar segmentation achieved Dice score of 0.959 ± 0.087 and Average surface distance of 0.873 ± 0.61 mm.