Editorial: Artificial Intelligence May Allow Reduced Use of Gadolinium-based Contrast Agents
Key findings
- One approach to reducing the risk of exposure to gadolinium-based contrast agents (GBCAs) is to reduce the dose, but the lower signal-to-noise ratio of the images acquired may not be adequate for detecting small tumors
- German researchers recently published a proof-of-concept study that suggests that generative adversarial networks (GANs) could be used to restore the signal-to-noise ratio of breast MRI scans acquired with lower doses of GBCAs
- In an accompanying editorial, Manisha Bahl, MD, MPH, of Massachusetts General Hospital explains how a GAN can predict the appearance of full-dose contrast-enhanced subtraction images, and she notes important strengths and limitations of the study
Several associated risks of gadolinium-based contrast agents (GBCAs) have been described in the medical literature and amplified in the lay press. The concerns include allergic-like hypersensitivity reactions, nephrogenic systemic fibrosis and, more recently, deposition of gadolinium in the brain, skin, bone and solid organs of patients with normal renal function.
Subscribe to the latest updates from Radiology Advances in Motion
One approach to reducing exposure to GBCAs is simply to reduce the dose. A prospective study of 104 women found that, for the detection and characterization of breast lesions with 3.0-T MRI, a 75% lower dose of gadobenate dimeglumine was noninferior to a dose of gadoterate meglumine that was higher than the on-label dose.
However, the lower signal-to-noise ratio of the images acquired with reduced-dose contrast material may not be adequate for the detection of small tumors in the screening setting.
German researchers recently published in Radiology a proof-of-concept study that describes how artificial intelligence could be used to restore the signal-to-noise ratio of breast MRI scans acquired with lower doses of GBCAs.
In an accompanying editorial, Manisha Bahl, MD, MPH, a radiologist in the Breast Imaging Division of the Department of Radiology at Massachusetts General Hospital, presents some caveats but agrees that the proof-of-concept study is an important step toward understanding how AI might aid in the quest to reduce the use of GBCAs.
Study Methods
The German researchers trained generative adversarial networks (GANs), which are deep learning–based models, to synthesize MRI scans that were similar in appearance to full-dose contrast-enhanced MRI scans. GANs are known for "deep fakes"—alteration or manipulation of images, recordings, or videos to misrepresent humans—but they can also be used to augment and synthesize medical data.
As Dr. Bahl explains, "One of the GAN's trainable networks, the generator, creates altered or manipulated images by adding noise. The GAN's other trainable network, the discriminator, attempts to distinguish the fake images from the real ones. A competition ensues between the generator and discriminator, both of which improve over time. Ideally, at some point, the generator produces images that are so realistic that the discriminator is unable to distinguish them from fake images."
The research team trained GANs with reduced-dose contrast-enhanced breast MRI scans and unenhanced MRI scans. The reduced-dose images were simulated (i.e., the signal-to-noise ratio of full-dose MRI scans was reduced to correspond to 25% of the contrast agent dose).
Then, those images and unenhanced T1- and T2-weighted images were inputted into the GANs, and the GANs predicted the appearance of full-dose contrast-enhanced subtraction images.
Key Study Results
When two blinded radiologists attempted to distinguish between the actual full-dose subtraction images and the GAN-generated subtraction images, their results were no better than random guesses.
When the radiologists were unblinded and compared the two sets of images, the GAN-generated images were assessed as noninferior to the actual images (average lesion conspicuity score of 4.9 on a five-point Likert scale).
Strengths and Limitations
Dr. Bahl notes that the German team used 9,551 breast MRI examinations for training, the largest attempt yet to use GANs to recover full-contrast breast MRI scans. Furthermore, the model was useful for detecting lesions as small as 5 mm and breast cancers manifesting as either a mass or nonmass enhancement.
The reason that the GANs weren't trained on actual reduced-dose MRI scans, Dr. Bahl points out, is that reduced-dose MRI scans are uncommonly obtained in clinical practice. That is an important limitation of the study. Moreover, all MRI scans used to train the GANs had a fixed noise level, which doesn't reflect real-world practice.
Next Steps
Issues to explore in future studies, Dr. Bahl says, are the effects of different levels of dose reduction, the sensitivity of the AI-generated images for the detection of breast tumors based on size and manifestation (mass vs. nonmass enhancement), and the diagnostic utility of AI-generated images in combination with diffusion-weighted imaging and other unenhanced sequences.
view original journal article Subscription may be required
Learn more about breast imaging at Mass General
Learn more about research in the Department of Radiology