Generative adversarial networks (GANs) offer potential in cross-modality image translation, but their application in pituitary adenomas remains uncertain. This study was to assess the feasibility of a GAN-based deep learning algorithm for generating synthetic diffusion-weighted imaging (DWI) associated images and its clinical utility in predicting tumor consistency. This multicenter study included a training cohort of 152 participants with large-to-giant pituitary adenomas from a tertiary center. Synthetic DWI associated images were generated from T2-weighted images (T2WI) and evaluated against real images. External validation was performed on a cohort of 69 participants from three additional centers to assess the utility of synthetic images in predicting tumor consistency and their correlation with surgical outcomes. In the independent test set, synthetic images demonstrated close resemblance to real images, with a mean squared error (MSE) below 70 and peak signal-to-noise ratio (PSNR) exceeding 30 dB. Neuroradiologists were largely unable to distinguish between real and synthetic images (P = 0.571) and rated comparable scores in the overall image quality (P = 0.051) and diagnostic confidence (P = 0.168) scale. Synthetic images demonstrate comparable performance to real images in predicting tumor consistency in both the independent test set (AUC = 0.84) and external cohort (AUC = 0.79). When combining features from all the three synthetic images, the performance in predicting tumor consistency outperformed conventional T2WI (P = 0.036 in the test set, P = 0.040 in the external cohort). Surgeons were more likely to adopt transcranial or combined approaches for predicted fibrous tumors, which were associated with lower tumor removal rates and higher severe complication rates compared to soft tumors. The GAN-based image synthesis method accurately predicts tumor consistency and shows potential for guiding surgical decision-making in patients with pituitary adenomas.
Wu et al. (Wed,) studied this question.