|
Generative adversarial networks (GANs) have been used to successfully translate images between multiple imaging modalities. While there is a significant amount of literature on the use cases for these approaches, there has been limited investigation into the optimal model design and evaluation criteria. In this paper, we demonstrated the performance of different approaches on the task of cone-beam computer tomography (CBCT) to fan-beam computer tomography (CT) translation. We examined the implications of choosing between 2D and 3D models, the size of 3D patches, and the integration of the Structural Similarity Index Measure (SSIM) into the cycle-consistency loss. Additionally, we introduced a partially-invertible VNet architecture into the RevGAN framework, enabling the use of 3D UNet-like architectures with minimal memory footprint. We compared image similarity metrics to visual inspection as an evaluation method for these models using held-out patient data and phantom scans to demonstrate their generalizability. Our findings suggest that 3D models, despite requiring a longer training time to converge due to the number of parameters, produce fewer image perturbations compared to 2D models. Training with larger patches also improved stability and significantly reduced artifacts, but increased the training time, while the SSIM-L1 cycle-consistency loss function enhanced performance. Interestingly, our study revealed a discrepancy between standard image similarity metrics and visual evaluation, with the former failing to adequately penalize visually evident artifacts in synthetic CT scans. This underscores the need for tailored and standardized evaluation metrics for medical image translation, which would facilitate more accurate comparisons across studies. To further the clinical applicability of image-to-image translation, we have open-sourced our methods and experiments, available at github.com/ganslate-team.
|
