We can create or reconstruct an image, such as that of a sphere or a triangle, by "connecting the dots". How many dots do we need to tell the difference between the two? Please refer to this figure. Much less if we have a certain template or model of the object i.e. a spherical or cubical model, and if AI can be used. Less dots or samples to collect means faster scans.
A challenge of image reconstruction was set by the Facebook Artificial Intelligence Research group and a medical MRI center. It consisted of accelerating MRI scans by sampling less points for image reconstruction and by using AI (e.g. neural networks) to connect the dots or fill the gaps. By sampling 4X and 8X less points (undersampling), the scan time was considerably shortened. It is concluded that there are minimal acquisition requirements, when neural networks fill the gaps.
Teams used convolutional neural networks (CNNs), recurrent neural networks (RNNs) and other model architectures.Â
"Results of the first fastMRI image reconstruction challenge"
AI Meta, 2019-01-12
"Facebook AI accelerates MRI exams"
Healthcare in Europe, 2021-01-21
"FastMRI breakthrough shows AI-accelerated MRIs interchangeable with traditional MRIs"
AI Meta, 2020-08-18