Bi-Directional MS Lesion Filling and Synthesis Using Denoising Diffusion Implicit Model-based Lesion Repainting

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Automatic magnetic resonance (MR) image processing is crucial for studying multiple sclerosis (MS).
• Lesions in MS complicate analyses like brain parcellation, and current lesion filling methods lack realism.
• Limited data with lesion delineations hinders the performance of lesion segmentation algorithms.

Purpose of the Study:

• To develop a novel approach for MS lesion filling and synthesis using Denoising Diffusion Implicit Models (DDIMs).
• To generate realistic lesion-free MR images and augment datasets for lesion segmentation tasks.

Main Methods:

• Leveraged Denoising Diffusion Implicit Models (DDIMs) for image inpainting.
• Modified DDIM architecture for both lesion filling and synthesis of lesioned tissues.
• Generated lesion-free T1-weighted or FLAIR images from lesioned ones, and synthesized lesions in healthy subjects' images.

Main Results:

• Demonstrated promising initial results in both lesion filling and synthesis tasks.
• The modified DDIM approach shows potential for creating realistic lesion-free images.
• The method can effectively augment training datasets for lesion segmentation algorithms.

Conclusions:

• The proposed DDIM-based approach offers a novel solution for MS lesion filling and synthesis.
• This technique can improve downstream analyses requiring lesion-free MR images.
• The synthesis capability is valuable for enhancing lesion segmentation model training.