• AI
  • Molecular Imaging
  • CT
  • X-Ray
  • Ultrasound
  • MRI
  • Facility Management
  • Mammography

Learning from Deep Learning in Radiology


There is a lot left to learn about deep learning.

Curtis P. Langlotz, MD, PhD, has been studying artificial intelligence since 1981, but admits there is still much to learn from computers.

You couldn’t turn a corner at RSNA 2017 without hearing about AI and how it can improve efficiency, accuracy, and usability of medical imaging.

“Up to 10% of patient deaths are related to some type of diagnostic error and 4% of radiology interpretations contain clinically significant errors,” Langlotz said. The promise of deep learning lies in helping reduce these errors.

Potential of Deep Learning in Radiology
There are many opportunities to use AI and deep learning in medical imaging: image quality control, imaging triage, efficient image creation, computer-aided detection, computer aided-classification, and automatic report drafting.

The research projects reported on by each speaker at the deep learning session consistently showed that deep learning outperformed humans in analyzing radiologic images. But in a world where everything is STAT, Luciano Prevedello, MD, from the Ohio State University said that deep learning can help in other ways aside from helping label images, such as improving process efficiency when dealing with many high priority cases. For example, his lab has been able to filter incoming images based on priority using deep learning. The algorithm looks at the images to identify brain hemorrhage or stroke, if the computer detects one of the flagged factors, the patient will move up on the priority list to have their images analyzed first. If the algorithm does not detect any critical factors, the patient’s case falls towards the bottom of the priority list.

There is a desire to have deep learning provide actionable advice that can be evaluated, which was emphasized by all three speakers. For example, at Stanford,  Langlotz described a deep learning algorithm that can improve MRI image quality and suggested a future where the MRI machine can notify the technologist that the images are too fuzzy to be read accurately. Through this type of approach, it is possible to improve MRI image quality  and have the patient spend less time in the machine.

What Does the Future Hold for Deep Learning?
“There is a hype cycle for emerging technologies-we are at the peak for inflated expectations about deep learning and machine learning, the trough of disillusionment is 2 to 5 years away. Some have predicted that radiologists will be replaced by robots, but because of the nature of a radiologists’ work, it is unlikely that a computer will be able to fully develop the complex analytic and reasoning skills required to completely replace human radiologists,” Langlotz said. For example, a human can understand when autopilot isn’t appropriate or doing the job well enough and you need to intervene, whereas a machine on autopilot will keep doing the same thing. The debate is still open on what level of impact deep learning and AI will ultimately have on radiologists and workflow in diagnostic imaging.

How Do We Handle the Reality of Deep Learning?
There are drawbacks to using deep learning algorithms, such as the single pixel error, where changing one pixel in an image that is being analyzed results in a different conclusion by the algorithm. “If the algorithms for deep learning are so sensitive, how do we know if the error moves the conclusion in the wrong direction?” Langlotz said Radiologists need to be careful about how they expand and use algorithms.

“It takes a village to get deep learning right and there are caveats of what to do with the data before running it through the algorithm. There is the challenge that the data are very messy-clinical data was not made for research-if the data is noisy can we still use this for clinical discoveries?” Prevedello said. “Does the data need to be curated all the time? Most of the time we do need to do some curation, deep learning should be able to figure out the noise.”

There is a lot of fear and excitement in the field about deep learning, but which should you be? Prevedello said, “This is a ride that everyone is getting to, so you might as well be excited.”

Related Videos
Can Diffusion Microstructural Imaging Provide Insights into Long Covid Beyond Conventional MRI?
Assessing the Impact of Radiology Workforce Shortages in Rural Communities
Emerging MRI and PET Research Reveals Link Between Visceral Abdominal Fat and Early Signs of Alzheimer’s Disease
Reimbursement Challenges in Radiology: An Interview with Richard Heller, MD
Nina Kottler, MD, MS
Related Content
© 2024 MJH Life Sciences

All rights reserved.