Temporal and spatially consistent inpainting with deep diffusion models

Most current brain imaging software is designed to work with images that have a consistent anatomy. However, this can be a problem when there are changes in the brain, like the appearance or disappearance of lesions, between two scans taken at different times. While we could develop complex new software for these situations, many researchers prefer to use existing, well-known tools, especially for studies over time. One suggested solution is to use a technique called image inpainting. In image inpainting, we fill in certain parts of an image based on a special pattern, making these parts look like they naturally belong there. This is really useful in brain scans, for example, when we need to add in images of lesions from multiple sclerosis (MS) in a way that looks real and fits with the rest of the brain image. Current inpainting methods usually practice on specific patterns, which means they might not work well on new or different types of patterns. Plus, these methods often struggle with handling lesions of various sizes and locations because they’re used to working with the same types of patterns, which limits how well they can work on different or new ones.

Faculty Supervisor:

Hassan Rivaz

Student:

Partner:

NeuroRx Solutions Inc

Discipline:

Computer science

Sector:

Health and Related Sciences & Technology; Professional, scientific and technical services

University:

Concordia University

Program: