The effect of weights for cycle-consistency loss and identity loss on blood-pool image to bone image translation with CycleGAN

Objectives: Patients must wait 2 or 4 hours after radioactive tracer injection for bone scintigraphy. This study aims to avoid the waiting by converting blood-pool images acquired 5 minutes after injection into bone images. Methods: Five hundred eleven patients with three-phase bone scintigraphy were enrolled. The blood-pool image was acquired 5 minutes after injection, and the bone image was obtained 3 hours after injection. We used 461 sets of the blood-pool images and the bone images as training datasets and 50 sets as testing datasets. A generative adversarial network with cycle-consistency loss and identity loss (CycleGAN) was used as a model for image-to-image translation. We tuned the weights of the cycle-consistency loss and the identity loss: Model A with 10 and 5 as the weights of the losses, Model B with 10 and 10, Model C with 100 and 50, and Model D with 100 and 100. We requested nuclear medicine radiologists perform the human eye perceptual evaluation (HYPE)-infinity method to evaluate the CycleGAN models' performances. Results: Nuclear medicine radiologists who participated in HYPE-infinity misjudged 4%, 0%, 30%, and 22% of each of the 50 fake images generated from Models A to D as real. And from Models A to D, the time taken to determine if it was fake or real was 8.21 ± 7.03, 9.88 ± 8.33, 11.95 ± 7.82, and 11.15 ± 9.09 seconds, respectively (p = 0.112). The 94% of real bone images were judged as real, and the time was 7.95 ± 4.03 seconds. Model C, which gave weights of 100 and 50 for the cycle-consistency loss and the identity loss, statistically showed the best performance among the models (p < 0.001).

Conclusions: Although not enough to fool the expert, more weights for the cycle-consistency loss and identity loss made the CycleGAN model translate the blood-pool image close to the real bone image. And the cycle-consistency loss could be more crucial than the identity loss.