One of the most transformative tools of medical imaging the last few decades has been magnetic resonance imaging, known to most as MRI. But, while powerful and insightful, MRI scans are a slow technology that can often lead to patient discomfort. For a patient already in distress, a typical scan lasting 40 to 60 minutes can be overwhelming.
Akshay Chaudhari, an assistant professor at Stanford University and a faculty member of the Institute for Human-Centered Artificial Intelligence (HAI), with collaborators Reinhard Heckel, an assistant professor at the Technical University of Munich, and Mohammad Zalbagi Darestani, a doctoral candidate at Rice University, recently explored how AI techniques could accelerate MRI. They found that AI techniques are at least as reliable as current, non-technical methods already in use and improve upon those techniques when clarifying small details in an image.
Read the full paper: "Robustness in Deep Learning Based Compressive Sensing"
Despite MRI being a workhorse for diagnostic radiology, Chaudhari says, its Achilles heel is that it just takes a long time. One way we can speed up scans is by collecting less data and using AI and other mathematical methods to reconstruct the images.
“Using AI for MRI reconstruction can have a multitude of practical benefits — patients can undergo much faster imaging procedures, the images have a lower likelihood of having artifacts due to patient motion, hospitals can cater to more patients with shorter wait times, and radiologists can still render accurate diagnoses for their patients,” Chaudhari says, listing the practical benefits of that speed.
The paper will be featured at the 2021 International Conference for Machine Learning beginning July 18, 2021.
Apples to Apples
MRI is useful for myriad purposes including scanning knees for cartilage and ligament damage or scanning the brain and body for tumors and cancers — diagnostics where false outcomes could have grave consequences. Recently, a team of mathematicians has even questioned the fundamental reliability of AI techniques in radiology, some saying they should not be used at all.
“Prior studies on the robustness of neural network-based image reconstruction drew an overly pessimistic picture of deep networks, but they only stressed the neural networks methods and not the classical ones already in use,” says Zalbagi Darestani.
To evaluate AI’s potential, the research team compared the two most promising AI techniques — trained and untrained neural networks — against non-AI-based image reconstruction methods currently in clinical use. Trained networks have been studied for the past few years and rely on high-quality examples to train against in a supervised manner. On the other hand, untrained networks represent cutting-edge advances in unsupervised AI that do not require any training data at all.
The researchers looked at how each model performed with regard to three common concerns affecting image reconstruction. The first involves sensitivity to worst-case disruptions in data collection — known as perturbations. Perturbations might be introduced by faulty MRI equipment, for instance, or by patients themselves when they move during the scan.
“Prior studies on the robustness of neural network-based image reconstruction evaluated perturbations intentionally tuned for neural network methods only, but not the state-of-the-art reconstruction methods,” Zalbagi Darestani notes. “It wasn’t an apples-to-apples comparison.”
The researchers also looked at how these models handle another form of data anomaly known as a distribution shift, which can occur in cases where an algorithm trained on one group of people is used on a new group or when scans trained of one part of the body, the knee for example, are used to reconstruct images of a different area of the body, say, the brain.
And, finally, Chaudhari and colleagues looked at the ability of each model to recover small details in an image — a nascent tumor, for instance.
The team’s most important finding was that all forms of image reconstruction — AI and traditional alike — are susceptible to these types of challenges. Therefore, Chaudhari says, the call to exclude AI methods, especially at this stage, is at best based on an incomplete or unfair comparison.
“Prior studies have only looked at perturbations designed for trained methods and applied them to traditional methods,” Chaudhari says. “In our study, we design worst-case perturbations separately for all three methods. We found that all image reconstruction systems are susceptible to the same concerns.”
“Perhaps the most surprising finding in our study is that neural networks prove no more sensitive to distribution shifts than un-trained methods, at least for MRI reconstruction,” Heckel adds.
In another notable outcome of the study, when it came to resolving fine detail however —where the consequences are most profound — the neural networks actually outperform traditional models.
The bottom line, Chaudhari says, is that neural networks are a promising and increasingly viable method for image reconstruction and should not be ruled out, particularly before they are fully evolved.
In fact, Chaudhari notes, given the similar susceptibility of all reconstruction methods to the same sorts of error, the group’s study should cement AI’s future in the field, particularly when weighing the modest risks of all forms of reconstruction against AI’s clear advantage in accelerating the MRI acquisitions. That performance advantage should equate directly with patient benefit through earlier detection and a reduction in anxiety-inducing false-positive diagnoses.
“The most important thing to keep in mind is what these risks and advantages mean for the patient. In that regard, neural networks are promising, and their capabilities should continue to grow with time,” Chaudhari concludes.
Stanford HAI's mission is to advance AI research, education, policy and practice to improve the human condition. Learn more.