Mayo Clinic researchers have been looking beyond the standard generative ML models for realistic medical imaging. They have come to introduce DDPMs (denoising diffusion probabilistic models), an ML-based diffusion technique to generate medical images. DDPMs are a relatively new class of generative ML models that enables the generation of labeled synthetic images.
Traditionally, generative ML models are used to learn from medical imaging data and generate realistic images that are not patient-specific. Researchers use these synthetic images to study medical conditions and abnormalities without compromising patient privacy. The only drawback is that such models generate unlabeled imaging data, which is selectively helpful in real-world applications.
However, the volume of medical imaging data with specific abnormalities is considerably less than that with common pathologies. This results in insufficiently large and imbalanced imaging datasets used for training the models, making them less accurate.
Diffusion models are based on the Markov chain theory and generate synthetic output by ‘gradual denoising’ of an image having Gaussian noise. The process is repeated for all images, making these models run significantly slower than others. Nevertheless, diffusion models outperform other generative models as they can extract more representative features from the input medical imaging data.
In proposing this research, the researchers created a tool that can retrieve 2D axial image slices from the FLAIR sequence of a brain MRI and inpaint a pre-defined area of that slice with a realistic image. The inpainted image can represent components, including the surrounding edema, tumors, or tumor-less brain tissues.
The diffusion model presented in ‘MULTITASK BRAIN TUMOR INPAINTING WITH DIFFUSION MODELS: A METHODOLOGICAL REPORT‘ will enable medical practitioners to induce/remove tumoral and tumor-less tissues using brain MRI slices with limited data. The researchers have also provided the code online for people to use.