SAN JOSE, Calif.--(BUSINESS WIRE)--At NVIDIA GTC, Inspur Information announced the results of its partnership with the Feinberg School of Medicine at Northwestern University in leveraging artificial intelligence (AI) to advance medical research and healthcare. Northwestern researchers have developed a custom AI workflow using Inspur AI servers with NVIDIA GPUs to accelerate the processing of radiology reports and provide crucial patient follow-up.

Medical diagnostic imaging from modalities such as X-rays, CTs and MRIs are reviewed, and findings are summarized in a radiology report which can contain recommendations for follow-up actions, such as further tests and evaluations. Due to the length and intricacy of these types of reports, up to 33% of follow-up recommendations are delayed or unintentionally overlooked, which can lead to poor patient outcomes. To solve this problem, Mozziyar Etemadi, MD, PhD, and his team at Northwestern developed an initiative to ensure reliable follow-ups of radiographic findings to prevent diagnostic and treatment delays and improve outcomes. The team developed an AI workflow based on recurrent neural networks and natural language processing (NLP) to examine and identify radiology reports with findings that require additional medical follow-up.

“We used AI and the tools at our disposal, including the Inspur NF5488M5-D GPU server featuring the NVIDIA A100 Tensor Core GPU,” said Dr. Etemadi. “We built our own custom AI workflow that reads nearly every single radiology report and, through deep integration with our medical record system, provides alerts and notifications to the primary care doctor, patient, and dedicated follow-up team, to ensure that important details do not fall through the cracks.”
In a study published in the New England Journal of Medicine Catalyst, Northwestern reported that its custom AI workflow screened over 570,000 imaging studies in 13 months and found 29,000—5.1% of the total—to contain lung-related follow-up recommendations, at an average rate of 70 findings flagged per day. Results demonstrated 77.1% sensitivity, 99.5% specificity, and 90.3% accuracy for follow-up on lung findings. Nearly 5,000 interactions with physicians were generated, and over 2,400 follow-ups were completed. The article concludes that AI and machine learning processes improve reliability of medical imaging findings, which can lead to effective reduction and prevention of high-risk diseases. The researchers have also released their open-source code with a tutorial at this link.

CT Homepage