Img-08. Synthesizing Missing Mri Sequences in Pediatric Brain Tumors Using Generative Adversarial Networks; Towards Improved Volumetric Tumor Assessment

Abstract BACKGROUND Standard MRI sequences, such as pre- and post-contrast T1w, T2w, and FLAIR images, are essential for optimizing segmentation of tumor subregions and evaluating treatment responses in pediatric brain tumors (PBTs). However, MRI sets are often incomplete due to imaging artifacts or inconsistent acquisition protocols across various centers. Generative Adversarial Networks (GANs) have been effectively utilized to generate missing MRI sequences for adult brain tumors. This study applies image-to-image translation models using GANs to synthesize missing FLAIR images from T2w images in PBTs. METHODS This retrospective study developed two GAN models, pix2pix in both 2D and 3D, trained and validated on T2w and FLAIR image pairs from 79 and 19 patients, respectively, with pediatric-type diffuse high-grade gliomas (diffuse midline glioma (DMG) including diffuse intrinsic pontine glioma (DIPG)), collected from the Children’s Brain Tumor Network (CBTN). The 2D pix2pix model processes each T2w MRI volume slice-by-slice, generating corresponding FLAIR slices to reconstruct the complete image volume. The 3D pix2pix model allows translation of 3D FLAIR volumes directly from T2w volumes without the need for individual slice processing and reconstruction. The quality of the generated FLAIR volumes was evaluated using Structural Similarity Index (SSI; best closer to 1) and Mean Squared Error (MSE; best near 0) on the validation set. RESULTS The 2D and 3D models achieved robust performance with median SSI, MSE of 0.88, 0.004, and 0.92, 0.002, respectively. CONCLUSIONS The GAN models developed in this study effectively generate missing FLAIR MRI volumes from corresponding T2w images in PBTs. The 3D model, outperforming 2D, speeds the overall processing pipeline by eliminating volume slicing and reconstruction. Future work includes assessing the impact of these synthesized images on the accuracy of our pretrained autosegmentation models in differentiating tumor subregions and measuring tumor volumes in longitudinal MRIs.