Deep Learning Shows Promise in Screening CT Scans
Healthcare applications for AI continue to emerge, including vetted algorithms used as diagnostic tools for analyzing medical imagery to spot critical abnormalities.
AI startup Qure.ai said a clinical validation study involving 21,000 patients found that its algorithm could help accurately identify internal bleeding, fractures or other trauma in head CT scans. It also made its data set of 491 CT scans analyzed by its algorithm available for download.
CT scans—computer-processed X-ray measurements taken from a variety of angles to produce cross-section images—are a standard diagnostic test for head injuries or stroke victims. Qure.ai, San Mateo, Calif., said its algorithm is intended to help fill the gap in reading scans. Radiologists may not be immediately available to interpret CT scans, and the startup noted that prompt assessment of CT scans is especially critical for stroke patients.
Prashant Warier, Qure.ai’s co-founder and CEO, claimed the algorithm could help screen CT scans in as little as 10 seconds to detect abnormalities and assess their severity.
The deep learning model was trained using more than 313,000 anonymized head CT scans. Of these, more than 21,000 were used to validate the AI algorithm. It was then clinically validated on 491 CT scans, with the results compared against a panel of three experienced radiologists that included a senior neuro-radiologist at the Mayo Clinic’s radiology department.
The deep learning system promises to automate as well as improve the quality and consistency of interpretation of head CT scans “as an adjunct to medical care,” said Dr. Norman Campeau, the Mayo Clinic neuro-radiologist who co-authored the Qure.ai study.
The validation study that found Qure.ai’s model was more than 95 percent accurate in identifying abnormalities was published here.
Qure.ai also has worked with GPU vendor Nvidia (NASDAQ: NVDA) on the CT scan effort, using its PyTorch deep learning framework along with Titan X and GeForce GTX platforms to train the model on the collection of labeled CT scans.
“Training deep learning models, especially in healthcare, is only one part of building a successful AI product,” Ankit Modi, a founding member of Qure.ai, added in a blog post. “There are other operational hurdles like convincing doctors to embrace AI” along with infrastructure challenges.
Automated interpretation of head CT scans is the latest example of how AI technology is being used to automate the interpretation of medical imagery, shortening the amount of time required for specialists to pinpoint stroke or other symptoms. In another example, Stanford University research Andrew Ng developed a neural network to automate detection of pneumonia in chest X-rays.
Founded in 2016 with funding supplied by Fractal Analytics, Qure.ai is among a handful of AI healthcare startups focused on deep learning systems. “Systems based on deep learning are already doing better than radiologists and existing algorithms in a variety of diagnostic tasks,” concludes a healthcare market survey released in January.