Figure 4.

Process of training and classification. After preprocessing and ROI cropping of nodule images, 2308 images were split into a training set and test set. Data augmentation based on histogram equalization/normalization and horizontal flipping increased the number of images fourfold. Transferred learning by three pretrained CNNs was performed with fivefold cross-validation; then, the diagnostic performance using the test set was evaluated. Region of interest (ROI), convolution neural network (CNN).