#356 – The Wire

  • The FDA’s AI Database: The FDA just published a complete list of FDA-cleared AI/ML-enabled medical devices, including 343 products overall and 241 radiology devices (yes, 70% of approved AI devices are for radiology). The FDA plans to continuously update this list, as part of its AI Action Plan.
  • Masks and Dictation Errors: Wearing face masks might increase dictation errors in radiology reports. That’s the unsurprising conclusion from a new UNC study that had six radiologists dictate 40 reports with and without masks, revealing a 25% higher error rate while masked (21.7 vs. 27.1 per 1k words). The good news is most errors were minor (11.9 vs. 15.2 per 1k), although masks also increased major error rates (5.6 vs. 7.3 per 1k).
  • GE Definium Tempo Unveiled: GE Healthcare announced its Definium Tempo fixed overhead tube suspension X-ray system, highlighted by a range of new technologist-friendly workflow and productivity features. The Definium Tempo’s RT-friendly features include a tube-mounted console (supports all exam setup and positioning at the system) and a suite of automated workflows intended to reduce exam times, errors, image variability, retakes, and physical strain.
  • End-to-End MRI AI: A Microsoft-led team developed an end-to-end deep learning framework that combines image reconstruction and pathology detection within a single cloud-based workflow (exams, pre-processing, DL reconstruction, DL lesion detection, image output), creating an automated process to evaluate DLIR models. The team trained and tested the framework to detect meniscus tears (typically a challenge with MRI DLIR) using knee MRIs captured at three acceleration rates, confirming that reconstructed MRI images still lack the fine image details needed for automatic DL-based pathology detection. However, this study is still a building block for future MRI DLIR research, and those future researchers can leverage the same fastMRI+ dataset used in this study.
  • Prestige’s PE: Prestige Medical Imaging (PMI) just took on a private equity investment from Atlantic Street Capital (ACS) that the major Eastern U.S. imaging dealer will use to fund its geographic and product portfolio expansions. PMI was already in growth mode, acquiring Southeast U.S. dealer G.E. Walker in February, while ACS has a history of scaling similar healthcare companies and plans to make “several” more medical imaging acquisitions going forward.
  • Bayer & Huma’s NSCLC Project: Bayer’s Oncology division and Huma are working together to develop a machine learning system that can distinguish different types of non-small-cell lung carcinomas (NSCLCs) in CT scans, allowing faster diagnosis and more personalized treatment. This is the latest step in the Bayer-Huma relationship, as Bayer contributed to Huma’s Series B (2019) and Series C (2021) rounds. It’s also the latest addition to the growing list of cancer treatment companies partnering with cancer detection companies to help catch more cancers while they’re treatable (and increase demand for cancer treatments).
  • CTC Evidence: CT colonography (CTC) could be the most effective non-invasive colorectal cancer (CRC) screening test, detecting more CRCs and avoiding more colonoscopies than multitarget stool DNA testing (mt-sDNA) and fecal immunochemical testing (FIT). An AJR study review (studies: 10 mt-sDNA, 27 CTC, and 88 FIT) found that when using a ≥10mm polyp size threshold, CTC had a 6.6% positive test rate (vs. 13.5% & 6.4%), a 6% CRC PPV (vs. 2.4% & 4.9%), a 4% advanced neoplasia detection rate (vs. 3.4% & 2%), and a 61% advanced neoplasia PPV (vs. 26.9% & 31.8%).
  • Leveljump’s Imaging Centers: Canadian teleradiology company Leveljump Healthcare now has a physical presence, acquiring three Ontario imaging centers for $4.3m. The new imaging centers will become part of Canadian Teleradiology Services (a Leveljump subsidiary), and although teleradiology remains Leveljump’s primary business, the company plans to continue to acquire imaging centers going forward.
  • Mutual Information AI: MIT researchers developed a new AI learning approach that analyzes both images and text in radiology reports and then leverages the image/text correlations within each report (aka mutual information) to train more accurate imaging AI models. The MIT mutual information model uses three CNNs (analysis of image portions, sentence-level text analysis, image/text mutual information analysis) to classify downstream images – potentially more accurately than AI models only trained with images. 
  • VA Whistleblower Punished: The Iowa City VA radiology department is on the wrong side of the news this week, amid reports that a veteran technologist was unfairly punished after reporting (and eventually testifying to Congress) that diagnostic exams were being cancelled without physician approval. Earlier this year, the technologist was ordered back to the same unit that he reported, but with a lower salary and less patient care responsibilities. He’s currently in mediation with the VA, and he’s making his concerns about VA whistleblower protections public.

You're signed up!

It's great to have you as a reader. Check your inbox for a welcome email.

-- The Imaging Wire team