Cercare Bulletin, Issue 6 from October 6, 2020
What you missed in AI radiology conferences; FDA’s new digital health center for radiology technology; first U.S. government reimbursement for AI radiology usage; teleradiology to grow to $22.8 billion
Welcome to our fourth monthly newsletter on neuroradiology and AI. We hope you find it helpful. Please feel free to share it with others.
If you would like to be informed when the new issue is out, please sign up here.
Conference Coverage: What are the killer apps for AI in MRI?
Several recent conferences covered AI issues in radiology.
Aunt Minnie’s fall virtual conference held six sessions on AI in imaging, ranging from challenges in integrating AI with enterprise imaging, to the future of AI in radiology. In the session on how AI can help radiology reinvent itself in a post-COVID-19 world, Dr. Eliot Siegel said the pandemic will have more of an impact on how diagnostic imaging is practiced than anything else has in the past 50 years. In a session on improving the perception about AI and radiology jobs, Dr. Sonia Gupta said it’s up to radiologists to shape the narrative. “If AI can improve radiologists’ efficiency and decrease burnout, that will translate to better patient care,” she said.
What will be the killer apps for AI in MRI? At a Society for MR Radiographers & Technologists (SMRT) meeting, Greg Zaharchuk, PhD, of Stanford University in California spoke about where AI could have a substantial clinical impact in MRI application. It is not image interpretation. He proposed that AI-based image reconstruction and postprocessing methods would be the first implemented, and image reconstruction the last. That’s because AI must address important clinical needs for current suboptimal processes. AI must also be robust and easy to implement and have commercial support. He said that good AI applications should focus on common but tedious clinical tasks like segmentation, as current methods are time-consuming and labor-intensive. In addition to image interpretation, less attractive AI applications focus on rare diseases, and tasks that are easy for radiologists.
Neuroradiology and AI in the news
FDA launches digital health center to coordinate policy and regulatory approaches for radiology technologies. The U.S. Food and Drug Administration (FDA) launched a digital health center within the Center for Devices and Radiological Health, to ensure rapid development and review of digital health technologies. The center will focus on artificial intelligence and machine learning in software as a medical device, cybersecurity, the Pre-Cert program, mobile health devices, wearables as a medical device, and technologies used to study medical products.
The U.S. approves reimbursement for radiology AI for the first time. The U.S. Centers for Medicare and Medicaid Services (CMS) approved reimbursement for a radiology AI application for the first time. The new technology add-on payment (NTAP) was given to Viz.ai for its ContaCT stroke detection algorithm in brain CT scans. Reimbursement will be up to $1,040 under specific conditions, in addition to the radiologist professional fee. This will help eliminate barriers to using AI technology.
China creates a national radiology image database. China created the country’s first national radiology image database to assist with sharing patient imaging information for AI. It will involve up to 400 hospitals that currently store images separately and share them only among affiliated hospitals or with agreements.
UK’s National Health Service funding AI and digital tools for imaging and pathology. The United Kingdom is funding the Imaging Artificial Intelligence Centres of Excellence and Digital Pathology with an additional £50 million to develop AI and digital tools for disease diagnoses. The goal is to develop more accurate diagnoses and personalized treatments, freeing up clinicians’ time and saving patients lives.
Comparison and validation of white matter hyperintensities segmentation software in elderly patients. A study in Neuroimage, Clinical compared automatic methods to segment white matter hyperintensities (WMH) in elderly populations to determine if they could help researchers and radiologists select the best method for application to use. They studied a research dataset with 147 patients, more than half from the Alzheimer’s Disease Neuroimaging Initiative, and the rest from a cohort referred for cognitive impairment. Researchers compared manual segmentation to automatic segmentation using FLAIR MRI, and seven methods producing segmentation mask which can be used by radiologists untrained in computer programming (methods included LGA, LPA, BIANCA, UBO detector, W2MHS, and nicMSlesion, with and without retraining). NicMSlesion, with retraining on the dataset, had the best results.
Detecting large vessel occlusion at multiphase CT angiography by using a deep convolutional neural network. A study published in Radiology showed that deep learning can help quickly detect arterial blockages causing strokes. The study involved creating a deep learning algorithm for multiphase CTA at Brown University in Rhode Island to recognize large vessel occlusions in CT angiography (CTA). They retrospectively tested the model on multiphase CTA examinations of 62 patients, and the model showed 100% sensitivity, detecting 31 large vessel occlusions. This was an improvement over a 77% sensitivity rate for single-phase CTA. Researchers say this is the first study to use multiphase CTA to look at occlusions in both the anterior and posterior head and neck. They will next test the algorithm prospectively.
Radiology’s use in telehealth will continue growing even after the pandemic ends. Radiologists are in demand for the growing geriatric population, and AI programs have the potential to eliminate tedious and time-consuming tasks like segmentation, giving radiologists more time to spend on other image reads and consultations. Integrating AI into the workflow and gaining acceptance still has challenges, but AI applications in neuroradiology continue to find both clinical need and applicability.