Figure 2

Mandibular tomography generated with different training times. The training time starts with the first batch of mandibular tomography images entering the learning stage. Since the discriminator neural network and the generator neural network are constantly changing during training, the generated image will change. The goal of the training is to obtain a discriminator that can tell whether a drawing looks similar to the mandible and a generator that can draw different tomographic images of the mandible. (a) Shows the loss value and smooth loss of the discriminator at different training times. A discriminator loss value closer to 0 indicates that the discriminator can distinguish the real image and the image generated by the generator. Since the change in loss value is drastic, we draw the smooth loss with a smoothness of 0.85 to facilitate the analysis. (b) Shows the loss value and smooth loss of the generator at different training times. A generator loss value closer to 0 means that the image generated by the generator is more likely to be treated as a real image by the discriminator, indicating a high degree of authenticity of the generated image. (c) Shows the images generated by the generator with the same input parameters at different training times. (d,e) Show the discriminator and generator adversarial training process. In process (e1), the loss of the generator decreases, indicating that the generator is making progress, which causes the loss of the discriminator to increase in process (d1); in other words, its capacity to judge truth and falsehood decreases. In process (d2), the loss of the discriminator decreases, indicating that the discriminator has improved and increasing the loss of the generator. Thus, the authenticity of the images generated by generator needs further improvement. More training details of the continuous process are provided in the article video (https://CTGANs.kuye.cn/).