Deep learning algorithms

We used two state-of-the-art deep learning algorithms, namely ResNET and CycleGAN models. The details of the deep learning approaches and architectures of the networks are presented in the “Supplementary information” section. The flowchart of CycleGAN architecture is presented in Fig. ​Fig.1.1. The motivation behind the choice of these architectures instead of popular networks, such as UNET, is that the adopted CycleGAN internally uses a UNET-like structure (generator based on UNET). In addition, although ResNET is a non-standard network for image synthesis, our trial and error experiments revealed that it outperforms UNET. The training and hyperparameter tuning of the models were performed on 60 patients. Ten patients were used for model evaluation whereas a separate unseen dataset of 15 patients served as a test (external validation) dataset. For data normalization, we converted PET images to standardized uptake values (SUVs) and then divided them by a SUV_max of 10.

Schematic architecture of the cycle-consistent generative adversarial network (CycleGAN) model used for FD PET synthesis. The left panel depicts the training process whereas the right panel shows the process of testing and the structure of the generator and discriminator

The deep learning models were implemented on NVIDIA 2080Ti GPU with 11 GB memory running under the windows 10 operating system. The training was performed using a mini-batch size of 6 for 215 epochs. We opted for not using cross-validation since recent guidelines seem to suggest that although multiple internal cross-validation can be useful, independent validation using an external dataset for a single trained model is preferred over internal validation to properly evaluate generalizability [25, 26].