Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust 3D Brain MRI Inpainting with Random Masking Augmentation

Juexin Zhang, Ying Weng, Ke Chen

TL;DR

The paper tackles pathology-induced bias in brain tumor MRI by introducing a 3D inpainting framework that synthesizes subject-specific healthy anatomy. It employs a 3D U-Net trained with a hybrid MAE+SSIM loss and a random masking augmentation strategy to produce high-fidelity healthy tissue in MRI scans, enabling controlled pathology transplantation for data augmentation. On BraTS-Inpainting 2025, the method achieves a final-test SSIM of 0.919, PSNR of 26.932, and RMSE of 0.052, securing first place and outperforming earlier winners. This approach provides realistic healthy proxies to counteract disease expression biases, supporting more robust and generalizable neuro-oncological AI models and facilitating counterfactual analyses for dataset expansion.

Abstract

The ASNR-MICCAI BraTS-Inpainting Challenge was established to mitigate dataset biases that limit deep learning models in the quantitative analysis of brain tumor MRI. This paper details our submission to the 2025 challenge, a novel deep learning framework for synthesizing healthy tissue in 3D scans. The core of our method is a U-Net architecture trained to inpaint synthetically corrupted regions, enhanced with a random masking augmentation strategy to improve generalization. Quantitative evaluation confirmed the efficacy of our approach, yielding an SSIM of 0.873$\pm$0.004, a PSNR of 24.996$\pm$4.694, and an MSE of 0.005$\pm$0.087 on the validation set. On the final online test set, our method achieved an SSIM of 0.919$\pm$0.088, a PSNR of 26.932$\pm$5.057, and an RMSE of 0.052$\pm$0.026. This performance secured first place in the BraTS-Inpainting 2025 challenge and surpassed the winning solutions from the 2023 and 2024 competitions on the official leaderboard.

Robust 3D Brain MRI Inpainting with Random Masking Augmentation

TL;DR

The paper tackles pathology-induced bias in brain tumor MRI by introducing a 3D inpainting framework that synthesizes subject-specific healthy anatomy. It employs a 3D U-Net trained with a hybrid MAE+SSIM loss and a random masking augmentation strategy to produce high-fidelity healthy tissue in MRI scans, enabling controlled pathology transplantation for data augmentation. On BraTS-Inpainting 2025, the method achieves a final-test SSIM of 0.919, PSNR of 26.932, and RMSE of 0.052, securing first place and outperforming earlier winners. This approach provides realistic healthy proxies to counteract disease expression biases, supporting more robust and generalizable neuro-oncological AI models and facilitating counterfactual analyses for dataset expansion.

Abstract

The ASNR-MICCAI BraTS-Inpainting Challenge was established to mitigate dataset biases that limit deep learning models in the quantitative analysis of brain tumor MRI. This paper details our submission to the 2025 challenge, a novel deep learning framework for synthesizing healthy tissue in 3D scans. The core of our method is a U-Net architecture trained to inpaint synthetically corrupted regions, enhanced with a random masking augmentation strategy to improve generalization. Quantitative evaluation confirmed the efficacy of our approach, yielding an SSIM of 0.8730.004, a PSNR of 24.9964.694, and an MSE of 0.0050.087 on the validation set. On the final online test set, our method achieved an SSIM of 0.9190.088, a PSNR of 26.9325.057, and an RMSE of 0.0520.026. This performance secured first place in the BraTS-Inpainting 2025 challenge and surpassed the winning solutions from the 2023 and 2024 competitions on the official leaderboard.

Paper Structure

This paper contains 13 sections, 1 equation, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Dataset
  4. Pre-processing
  5. Data Augmentation
  6. Network Architecture
  7. Loss Function
  8. Experiment Results
  9. Evaluation Metrics
  10. Experiment Settings
  11. Validation Phase
  12. Test Phase
  13. Conclusion

Figures (3)

  • Figure 1: The figure illustrates the architecture of our U-Net model.
  • Figure 2: Qualitative results of our model's infilling performance on validation MRI scans, showcasing the best (Fig \ref{['sub1']}), median (Fig \ref{['sub2']}), and worst (Fig \ref{['sub3']}) cases. The green masks indicate the inpainted regions, which contained both healthy and unhealthy tissues as these were not explicitly labeled.
  • Figure 3: Qualitative comparison of our method's worst-performing case against the winning methods of the BraTS 2023 and 2024 challenges.