Fig. 2

Overview of Deep learning workflow. The workflow of both the training and test processes was presented. In the training session, image processing was performed as follows: 1) slice selection; the specific slice which included the largest tumor area was selected, 2) grayscale adjustment; setting the pixel with its SUV of 0 (=lower limit) and its SUV of 30 (=upper limit), 3) image conversion from the DICOM to JPEG picture data, 4) data augmentation; an additional 23 images were generated from each image. These images were fed into the deep learning training session, thereafter, the axial and coronal image-based diagnostic model was respectively created. In the test session, we performed the same imaging processing without data augmentation in original images of test cohorts. These processed images were classified by the diagnostic model created in the training session. Finally, diagnostic performance by axial, coronal, and axial-coronal combination use was respectively obtained