Anatomy-Aware 3D GAN with Local Mask Refinement for Brain MRI Inpainting
Brain tumor MRI inpainting aims to synthesize plausible healthy-appearing tissue in regions removed from pathological scans. We present an anatomy-aware three-dimensional generative adversarial network for Task 4 of the BraTS 2026 challenge. The method combines mask-aware gated convolutions, mirrored contralateral context, attention-modulated skip fusion, boundary-aware refinement, and a local masked-region refiner, each adapted from established components in the image synthesis and inpainting literature. The model is trained with a two-stage objective that first emphasizes voxel-wise and structural reconstruction and then applies low-weight adversarial fine-tuning. On a deterministic local held-out split, the proposed method substantially improves over a simple 3D AutoEncoder baseline but does not surpass the challenge's Pix2Pix3D baseline. On the official BraTS 2026 Task 4 evaluation server, the submitted model obtained a mean SSIM of 0.658 and a mean PSNR of 14.81~dB over 219 evaluated cases, lower than the local held-out figures. We report these results transparently, analyze likely contributing factors, and identify the controlled ablations (removal of the mirrored-context channel and of the two-stage schedule) required to isolate the contribution of each architectural component.
Authors
- Hari Mohan Rai (ORCID: https://orcid.org/0000-0003-2557-3510)
- Sayed Yasar Ahmad Mushtaq (ORCID: https://orcid.org/0009-0002-4828-7007)
- Abdul Fatah Rahimi (ORCID: https://orcid.org/0009-0004-8970-4422)
Institutions
- Nazarbayev University (KZ)
Publication Details
- Journal
- medRxiv
- Published
- 2026-11-15
- DOI
- https://doi.org/10.64898/2026.11.12.26367904
- Primary Topic
- Generative Adversarial Networks and Image Synthesis
- Type
- preprint