MMIT-DDPM - Multilateral medical image translation with class and structure supervised diffusion-based model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Unified medical image translation across diverse modalities is crucial for healthcare.
• Existing one-to-one Generative Adversarial Networks (GANs) struggle with multi-modal translation diversity.
• Denoising Diffusion Probabilistic Models (DDPMs) show promise in image generation due to scalability and data distribution capture.

Purpose of the Study:

• To propose a novel conditioning mechanism for deterministic, one-to-many medical scan translation using a DDPM.
• To enable unified translation among T1-weighted, T2-weighted, and FLAIR MRI sequences.
• To eliminate the need for multiple specialized bilateral translation models.

Main Methods:

• A novel conditioning mechanism was developed for DDPM-based medical image translation.
• The model utilizes a source-modality structure encoder and a source-to-target class conditioner.
• Training and testing were performed on the BraTS 2021 dataset, including T1ce, T2, and FLAIR sequences.

Main Results:

• The proposed DDPM achieved unified multi-lateral translation across six combinations of brain MRI sequences.
• The diffusion model demonstrated superior performance over state-of-the-art GANs, including Convolutional Neural Networks (CNNs) and Transformers.
• Quantitative improvements included an 8.06% increase in Multi-Scale Structural Similarity (MSSIM) and a 2.52 decrease in Fréchet Inception Distance (FID) metrics.

Conclusions:

• The developed DDPM with a novel conditioning mechanism enables efficient and unified one-to-many medical image translation.
• This approach surpasses traditional GANs in image quality and modality coverage for brain MRI.
• The findings pave the way for more versatile and effective medical image processing tools.