Abstract
Healthy tissue inpainting has significant applications, including thegeneration of pseudo-healthy baselines for tumor growth models and thefacilitation of image registration. In previous editions of the BraTS LocalSynthesis of Healthy Brain Tissue via Inpainting Challenge, denoising diffusionprobabilistic models (DDPMs) demonstrated qualitatively convincing results butsuffered from low sampling speed. To mitigate this limitation, we adapted a 2Dimage generation approach, combining DDPMs with generative adversarial networks(GANs) and employing a variance-preserving noise schedule, for the task of 3Dinpainting. Our experiments showed that the variance-preserving noise scheduleand the selected reconstruction losses can be effectively utilized forhigh-quality 3D inpainting in a few time steps without requiring adversarialtraining. We applied our findings to a different architecture, a 3D waveletdiffusion model (WDM3D) that does not include a GAN component. The resultingmodel, denoted as fastWDM3D, obtained a SSIM of 0.8571, a MSE of 0.0079, and aPSNR of 22.26 on the BraTS inpainting test set. Remarkably, it achieved thesescores using only two time steps, completing the 3D inpainting process in 1.81s per image. When compared to other DDPMs used for healthy brain tissueinpainting, our model is up to 800 x faster while still achieving superiorperformance metrics. Our proposed method, fastWDM3D, represents a promisingapproach for fast and accurate healthy tissue inpainting. Our code is availableat https://github.com/AliciaDurrer/fastWDM3D.