Denoising Diffusion Models for 3D Healthy Brain Tissue Inpainting

Automated analysis of magnetic resonance images, e.g., by measuring volumetric changes, is essential for monitoring diseases impacting the brain's structural integrity. However, many existing evaluation tools are designed primarily for healthy tissue analysis. To facilitate the assessment of scans with pathological tissue, it is necessary to restore healthy tissue in affected areas. In this study, we explore and extend denoising diffusion probabilistic models (DDPMs) for the inpainting of 3D consistent healthy brain tissue.


We enhance state-of-the-art 2D, pseudo-3D, and 3D denoising diffusion probabilistic models (DDPMs) that operate in image space, as well as 3D latent and 3D wavelet DDPMs, allowing them to synthesize healthy brain tissue. Our evaluation reveals that the pseudo-3D model achieves the best performance in terms of structural similarity index, peak signal-to-noise ratio, and mean squared error. To highlight its clinical significance, we fine-tune this model on synthetic multiple sclerosis lesions and assess its performance on a downstream brain tissue segmentation task, where it surpasses the established FMRIB Software Library (FSL) lesion-filling method. More details can be found in [1].

Slice Inpainting

Exemplary axial and sagittal slices of the same subject. The DDM pseudo3D performs best amongst all evaluated methods. In contrast to, e.g., DDM 2D slice-wise, which produces stripe artifacts due to the stacking of the generated 2D slices (blue rectangle), it shows 3D consistent inpainting.

Lesion Filling

The FSL lesion filling produces dot-like artifacts, which are not present in the inpainting generated by the DDM pseudo3D.

To top