Dr. Luciano Prevedello

The promise of AI (part 2 of 2)

September 28, 2018
by Sean Ruck, Contributing Editor
In last month’s HCB News, we spoke with Dr. Luciano Prevedello, a radiologist at the Ohio State University Wexner Medical Center and chief of the division of imaging informatics, about the past, present and future of artificial intelligence in healthcare, and the potential benefits (and challenges) it brings to medical imaging.In part two, Dr. Prevedello shares insight he’s gained through research and his own experience at the AI lab his radiology department created.

To start with, we talked about AI’s implementation. Prevedello believes that at least in the beginning, AI will need some degree of local development and/or validation prior to full implementation. New algorithms like deep learning have been performing extremely well, but tend to require large amounts of data for training and validation. Given that access to medical images is governed by several privacy rules, image sharing and algorithm development can only happen with the appropriate partnerships and agreements in place or at a site that has both access to the data and machine learning expertise. “Due to these limitations, researchers have been experimenting with different solutions. Instead of having data leave the institution to train algorithms at specialized facilities, one of the ideas is to have the algorithms come to the institutions. It is possible that this will become a trend in the future – the distribution of the platforms and algorithms to the institutions rather than data leaving the institution, but it will take time,” Prevedello said.

The investment of time and money by AI pioneers looks like it will pay off, as even today, the technology is showing fairly good accuracy at classifying or identifying images of interest for radiologists. Prevedello cautioned, however, that it’s still far from perfect. He suggests that for now, one way to use these tools is to help improve workflow by pre-screening studies for potential critical findings thereby expediting diagnosis and treatment decisions. Prevedello warns against the indiscriminate use of these tools and believes that extensive validation is needed prior to implementation to assist with image interpretation. “First, we need to understand a lot more about the tools – when they fail and why. There’s still a lot of research that needs to happen,” he said.

Taking a look behind the scenes of what makes AI work gets technical, but it’s still digestible concepts at the surface level. Recent image classification tasks have been employing convolutional neural networks (CNN) to analyze visual information from either 2D or 3D datasets. One of the challenges is that there is extensive data variety and complexity in medical imaging. For example, while some modalities can produce color images, as one might see in some Doppler ultrasound exams, or in PET/CT, other technologies, like X-ray or MR generate grayscale images. These modalities also have different resolutions and different ways to represent specific body tissues. New algorithms need to account for these differences. “Many of the algorithms being used now were created based on a dataset called ImageNet,” said Prevedello. A challenge was created around ImageNet to create algorithms that could identify specific objects within images using AI. “ImageNet used photographic color images – an airplane, apple, oranges – the task was to classify images into one thousand object classes. The winning algorithms became very well-known and they’re used for multiple purposes today. In medical imaging, we use these algorithms as well. While they were tailored for color images, we learned how to reconfigure them for grayscale,” he said.

There are additional ways that the information gathered from medical images differ from standard photographic images. Prevedello discussed CT further, where an exam will include a range of numbers. “So 0 is water, 2000 would be bone, -1000 would be air, and -75 for example would be fat . . . different values represent different tissue types, and this is extremely important for lesion characterization. Same thing with MRI, just different process. You have a high signal on T1, it can be blood, it can be fat – the signal characteristic means something. It has a tissue characterization embedded into them. So the way you deal with these images on the pre-processing side is different than a photographic image because you don’t want to lose that type of information, which is key for diagnosis, patient follow-up and the complete understanding of the clinical picture,” Prevedello explained.

The increase in use of AI going forward will undoubtedly impact the work of radiologists, but a lot remains to be seen on exactly how it will impact the work. For now, Prevedello stresses that it’s important for radiologists to at least understand the general concepts of how algorithms are built and how they might fail. That’s why he’s a proponent for educating residents and medical students, because as he pointed out, “you only identify things you know, and if you don’t know about a particular issue an algorithm may have in detecting or classifying something, you may not interpret that information correctly.” Meaning at least for now, artificial intelligence is still no substitute for human intelligence.