Sure -- i guess the idea/hope here is that the information itself might actually be in the 3T image, but in such a manner that is not perceptible by humans, or quantifiable using current approaches. We've already seen some evidence of that, but we def need more clinical data to figure it out

Many included participants had fairly extensive brain atrophy due to old age and/or neurodegenerative disease. Or did you mean something else?

Led by brilliant MSc student Malo Gicquel and co-supervised by equally brilliant Gabrielle Flood. We also got great contributions from @anikawuestefeld.bsky.social @xiaoyucaly.bsky.social @rikossenkoppele.bsky.social @lemwisse.bsky.social @biofinder.bsky.social @davidberron.bsky.social
and others!

We are working on assessing utility of this 3T->7T model in clinical situations, improving it, making it more generalizable, and extending it to other sequences. Please get in contact if you have some paired 3T-7T data, even if its a small dataset!

This is super important — it means increasing the quality of the image is not just eye candy. It might actually improve some of the downstream tasks we use the MRI for, like segmentation and visual reads.

Most impressively, when we ran automated amygdala segmentation on the real 3T and synthetic 7T, the segmentations from the synthetic 7T better matched manual segmentations of the amygdala!!

Our models outperform existing models on this task in terms of traditional metrics. However, a set of four blinded radiologists and MRI professionals also subjectively rated the synthetic images to be of higher visual quality than the images they were synthesized from.

Shockingly, when applied to unseen 3T scans, our models produce synthetic scans that resemble the original images, but sharper and with better contrast, while avoiding common artifacts seen on 7T scans.

We exploited a valuable dataset of 172 cognitively impaired and unimpaired older people from
@biofinder.bsky.social with paired 3T and a 7T T1-weighted scans. To this dataset, we apply our own specialized U-Net and GAN U-Net model, as well as some previously described models.

To improve accessibility of these high quality scans, we look to synthetic AI “super-resolution”, which has already been successful in brining lower field (0.5T - 1.5T) to 3T resolution.

7 Tesla (7T) MRI is important for studying fine anatomical morphology, and has found important applications in clinical research centers focusing on epilepsy and MS. But, its ceiling is underexplored; <150 7T scanners exist worldwide!

7T world map: google.com/maps/d/u/0/v...