Region-guided focal adversarial learning for CT-to-MRI translation: A proof-of-concept and validation study in...

Generative adversarial networks (GANs) have recently demonstrated significant potential for producing virtual images with the same characteristics as real-life landscapes, thereby enhancing various medical tasks.

To design a region-guided focal GAN (Focal-GAN) for translating images between CT and MRI and test its clinical applicability in patients with hepatocellular carcinoma (HCC).

Between January 2012 and October 2021, two cohorts of patients with HCC who underwent contrast-enhanced CT (Center 1, n = 685) and MRI (Center 1, n = 516; Center 2, n = 318) were retrospectively enrolled. We trained the Focal-GAN model by adding tumor regions to a baseline Cycle-GAN framework to steer the model toward focal attention learning. The quality of the images generated was assessed using an open-source MRQy tool. The clinical applicability of the Focal-GAN was evaluated by applying the nnUNet and ResNet-50 model for tumor segmentation and microvascular invasion (MVI) prediction in HCC on the generated images.

In the ablation tests, Focal-GAN achieved a higher fidelity than the conventional Cycle-GAN in the generated image quality assessment with MRQy. Regarding applicability, regardless of tumor size, nnUNet trained with focal-GAN-generated images achieved higher Dice scores than nnUNet trained using Cycle-GAN-generated images for HCC segmentation in both internal (0.607 vs. 0.341, p < 0.01) and external (0.796 vs. 0.753, p < 0.001) validation. Additionally, ResNet-50 trained with Focal-GAN-generated images produced higher areas-under-curve (AUCs) than ResNet-50 trained with real images for MVI prediction in both internal (0.754 vs. 0.665, p = 0.048) and external (0.670 vs. 0.579, p < 0.001) validation.

The designed Focal-GAN model can generate virtual MR images from unpaired CT images, thereby extending the clinical applicability of CT in the liver tumor diagnostic pathway.