Fig. 2
From: Three-stage segmentation of lung region from CT images using deep neural networks
Flow chart for the generation of training data for CNN-1 and CNN-2 models. Three databases LIDR, PHTM and ILD were visually inspected (VS1, VS2, VS3) and three sets of grayscale images (TR1, TR2, TR3) was selected. These images are combined to into a single set and visually inspected (VS4) again to categorize them into images containing lung region CN1-L1 and images without lung region CN1-NL1. Otsu thresholding OTX is applied to images in each category which converts them to binary images and tagged training data, CN1-L and CN1-NL for CNN-1. The generation of training data for CNN-2 begins with applying k-means KMX clustering to the sets of images (TR1, TR2 and TR3) from the three databases. Application of connected component analysis CCX on the output TR4 of the clustering algorithm produces set of images TR5 which are visually inspected VS5 and categorized into components containing lung region CN2-L and components without lung region CN2-NL