AI rapidly produces higher quality medical imaging from less data
Researchers at Massachusetts General Hospital have developed a new medical imaging technique based on artificial intelligence designed to enable clinicians to acquire higher quality images without having to collect additional data.
The AI technique—called AUTOMAP (automated transform by manifold approximation)—produces high-quality images in less time with MRI or with lower radiation doses with X-ray, CT and PET. And, as a result of its very quick processing speed, the approach could help in making real-time clinical decisions about imaging protocols while the patient is in the scanner, according to MGH researchers.
A description of the technique, published last week in the journal Nature, shows dramatic differences between images reconstructed from the same data with conventional approaches compared to AUTOMAP.
“What we did was condition a neural network through machine learning to recognize what makes an image an image,” says Matthew Rosen, director of the Low Field MRI and Hyperpolarized Media Laboratory and co-director of the Center for Machine Learning at MGH’s Athinoula A. Martinos Center for Biomedical Imaging. “What the network is learning are generic properties of images, not detailed properties of normal or diseased pathology.”
Rosen notes that the speed of image reconstruction from the machine learning algorithm is nearly instantaneous, taking just tens of milliseconds—a significant advancement for medical imaging.
“Some types of scans currently require time-consuming computational processing to reconstruct the images,” he observes. “In those cases, immediate feedback is not available during initial imaging, and a repeat study may be required to better identify a suspected abnormality. AUTOMAP would provide instant image reconstruction to inform the decision-making process during scanning and could prevent the need for additional visits.”
According to Bo Zhu, a research fellow in the MGH Martinos Center and first author of the paper, image reconstruction transforms raw data from a scanner into images for radiologists to evaluate.
“The conventional approach to image reconstruction uses a chain of handcrafted signal processing modules that require expert manual parameter tuning and often are unable to handle imperfections of the raw data, such as noise,” says Zhu. “We introduce a new paradigm in which the correct image reconstruction algorithm is automatically determined by deep learning artificial intelligence.”
While image reconstruction requires an enormous amount of computation, especially during the training of the algorithms, Rosen contends that once trained AUTOMAP can work together with inexpensive graphical processing unit (GPUs)-accelerated computers to improve clinical imaging and outcomes.
“We’re very lucky here because my lab is part of Mass General Hospital and we have a Clinical Data Science Center which enables us to push our algorithms to the terminals of the radiologists,” he concludes. “In the next 6 months, maybe, we’ll start doing that. We’ll have some studies where our reconstructions are running side-by-side with the state-of-the-art, FDA-approved algorithms and let radiologists give us feedback on what they think.”