Title:	Super-resolution neuroimages from widely accessible clinical scans
DNr:	Berzelius-2024-156
Project Type:	LiU Berzelius
Principal Investigator:	Jacob Vogel <jacob.vogel@med.lu.se>
Affiliation:	Lunds universitet
Duration:	2024-06-10 – 2025-01-01
Classification:	30105
Keywords:

Abstract

Integration of artificial intelligence in medical image analysis has already had a large impact on e.g. diagnostics of various diseases, automatic segmentation, image processing… Recently, some models, mainly based on CNNs, GANs and diffusion models, have shown to be especially effective to improve the quality of natural and medical images, by enhancing their resolution (super resolution). Magnetic resonance imaging (MRI) is a medical imaging technique that takes in vivo 3D pictures of the body. It is widely used in clinics and hospitals and is especially useful for brain analysis. Indeed, it indicates gross (e.g. visible to the naked eye) issues with the brains, including tumors, strokes, and severe neurodegeneration. Moreover, we can get finer-grained information about the brain parts that are unhealthy with analysis techniques, which can help us diagnose patients and predict patient outcomes. The accuracy of the assessments and analysis heavily depends on the quality of the scans, which is mainly indicated by their resolution. Their resolution is measured in Teslas (T) i.e the power of the magnetic field used. Most clinical scanners will be 1.5T and in low-income countries, as low as 0.5T. Most research scanners nowadays are 3T, though they are occasionally used in wealthy clinics and specialized centers. 7T images are also used but for research only. The higher the resolution, the finer the view of brain structures. With 7T, we can even see laminar structures in the cortex, and can better image subcortical nuclei. Yet, these scans are rare, the data is hard to process, expensive, etc. BioFINDER has a set of about 250 paired 3T and 7T scans of their patients, as well as thousands of unpaired 3T scans. The idea is to train a deep learning model on this set to bring 3T images all the way to 7T resolution, this is called "super resolution". Lately, new super resolution models have proven to be especially effective. Thus, similar ideas have recently been explored, but mainly to bring 1.5T images to 3T resolution. The most recent models are yet to be used on 3T and 7T images, as only CNN models were used. We want to adapt these models to our data and use it for different purposes, starting with a CNN then a GAN and if we have the time a diffusion model. There are three main purposes : improving the visual assessment done by radiologists, improving the automatic diagnostics (ex : atrophy detection) and improving the visualization and segmentation of very small parts of the brain such as subcortical nuclei, heavily involved in neurodegenerative diseases. This project will uniquely use AI to gain insight and potentially clinical utility from large and rare human datasets. However, this project can only be executed using extensive computational and memory resources, especially as 7T images are very large, having as many as 26 million voxels. We plan on building our models on python, mostly using PyTorch.