Achieving Over 10 \(\times \) Faster Sample Generation with Conditional Denoising Diffusion
摘要
This paper presents our solutions for Task 8 and 9 of BraTS 2025, which respectively involve the generation of a missing MRI modality and inpainting a missing region. Task 8 aims to solve the problem of missing a MRI modality for cases which acquisition is infeasible or the quality is not good enough for analysis. Task 9 seeks to produce pathology-free cases which would allow the analysis of healthy brains. We use denoising diffusion models to solve both tasks in a unified framework. Since regular diffusion models require 1000 steps or more for inference, we propose a solution to speed up inference while obtaining competitive results while keeping a low computational footprint. Compared to our solution from BraTS 2024, our new solution achieves better results with over 10 times faster processing. We obtain Dice scores of 0.80, 0.83, and 0.88 for ET, TC, and WT, respectively, along with an SSIM of 0.95 on Task 8 test set. On the Task 9 test set, we achieve an RMSE of 0.053, a PSNR of 26.77, and an SSIM of 0.918. The code will be released following the conclusion of the challenge.