Convolutional neural networks (CNNs) can be trained to analyze visual imagery much easier than other artificial neural networks, making their capabilities especially important to the future of radiology, according to a new analysis published in Current Problems in Diagnostic Radiology.
“A major strength of the CNN architecture is that no explicit feature identification is required,” wrote authors Rebecca J. Mieloszyk, PhD, Puneet Bhargava, MD, of the department of radiology at the University of Washington School of Medicine in Seattle. “We need not specify that houses have straight edges and sloped rooves in order for a CNN to learn to recognize them. Rather than defining specific features or rules for image classification, CNN design decisions include sizing of the input image, depth of the network, and the classification loss function. These are much more readily tuned than hand-crafted feature vectors.”
CNNs have already been trained to identify objects such as cats, dogs and houses in images, as seen in the way smartphone users can successfully search their photo library for those terms. There are still occasional hiccups, the authors add, pointing to a photo of a yak that was incorrectly labeled as “dog,” but the technology has still been found to be incredibly accurate overall.
In radiology, of course, researchers see huge potential in these technologies—and for good reason. “Projects have demonstrated CNN use in image acquisition, segmentation, nodule detection, captioning, and classification tasks,” Mieloszyk and Bhargava wrote. “CNN-based tools have been used to reliably reconstruct magnetic resonance images from k-space. Automated segmentation methods using CNNs have also been investigated in several radiology settings. These approaches represent a significant improvement over the typically supervised atlases or rule-based guidelines necessary to train machine-assisted medical image segmentation tools.”
There are limits to the power of CNNs, however. The authors explained that, sometimes, researchers can’t quite figure out why a CNN makes a certain designation. In addition, the amount of hardware and data required for these networks to work is significant, potentially limiting their availability in “many clinical settings.” These limitations are understood by the entire industry, though, and artificial intelligence and radiology specialists alike are hard at work to find solutions.
Overall, Mieloszyk and Bhargava concluded, CNNs “will likely find traction as useful tools” in radiology, performing certain tasks while allowing radiologists to spend more time providing value-based care.
“Next time you tag an image in Facebook or use a keyword to search Google Photos, think of ways that the same technology might help you cluster similar imaging cases, track subtle changes in longitudinal studies, or quickly present to you key images for abnormalities,” Mieloszyk and Bhargava wrote. “The future clearly is exciting!”