Abstract Background Assessment of treatment response in pediatric diffuse midline gliomas (DMGs), aggressive tumors with high mortality, relies heavily on radiological tumor size measures, highlighting the need for reproducible approaches such as autosegmentation methods. Emerging evidence suggests that volumetric tumor measurements from longitudinal MRI better capture tumor shape irregularities compared to traditional two-dimensional measures. However, incomplete MRI sets, caused by artifacts or protocol inconsistencies, hinder automated volumetric approaches for predicting tumor growth trajectories. This study leverages Generative Adversarial Networks (GANs) to synthesize missing FLAIR and T1w-pre images from T2w and T1w-post images, respectively, enabling longitudinal tumor volume assessment with external multi-institutional clinical trial data. Methods Two GAN models were trained on FLAIR/T2w and T1w-post/T1w-pre pairs using a cohort of 475 pediatric brain tumors (PBTs) from the Children’s Brain Tumor Network (CBTN). Generalization was tested on an external cohort of 32 DMG patients with 4 longitudinal imaging timepoints from PNOC003 and PNOC007 clinical trials and CHOP. Median Dice scores and %error in tumor volume were calculated when replacing real FLAIR and T1w-pre sequences with (a) GAN-synthesized images, (b) T2w/T1w-post images, and (c) zero-padded images. Results GAN synthesis achieved a median %error/Dice of 16.85%/0.83, outperforming 40.79%/0.64 (existing scans) and 60.83%/0.27 (zero-padded images). Differences in %error and Dice between GAN synthesis, existing scans, and zero padded images were statistically significant (p 0.05) using the Friedman and Dunn tests. Conclusion GANs trained on CBTN data generalize well to longitudinal clinical data. Synthetic image imputation reduces %error in tumor volume estimation, advancing longitudinal volumetric response assessment in real-world clinical settings.
Chrysochoou et al. (Fri,) studied this question.