Algorithm flags 'uninformative' images that won't support diagnosis, AI training

por John R. Fischer, Senior Reporter | October 02, 2020
Artificial Intelligence X-Ray
Presagen's new AI algorithm, UDC, can detect errors in data from medical images that make it difficult for radiologists or AI solutions to interpret diagnoses
AI healthcare company Presagen has designed an AI algorithm that can automatically detect errors that make medical images difficult for either radiologists or AI solutions to effectively diagnose.

Known as UDC, the algorithm can potentially triage medical scans and identify cases that require further in-depth clinical evaluation or additional tests to verify a definitive clinical diagnosis.

"The types of errors that are difficult to identify by radiologists would just be missed in their diagnoses — they wouldn't necessarily know that the data has errors or is poor quality," Dr. Michelle Perugini, co-founder and CEO, told HCB News. "However, from an AI development perspective, those errors can trick the training of the AI and make it less accurate than if the error data was removed. The UDC allows removal of the data with errors, that makes the AI training process better and creates a more accurate and commercially scalable AI that can assess/diagnose radiology images."
DOTmed text ad

Reveal Mobi Pro now available for sale in the US

Reveal Mobi Pro integrates the Reveal 35C detector with SpectralDR technology into a modern mobile X-ray solution. Mobi Pro allows for simultaneous acquisition of conventional & dual-energy images with a single exposure. Contact us for a demo at no cost.

Presagen used UDC to interpret a series of X-ray images for the detection of pneumonia. While errors were rare among radiologists when images had clear features, the algorithm found the diagnosis or label for several X-rays to be neither correct nor wrong. An independent radiologist who verified the images agreed they were difficult scans to diagnose, in contrast with the original diagnosis made in the public data set. AI that was trained to diagnose pneumonia also found the assessment for these images difficult.

Removal of these images from the training dataset improved AI accuracy for diagnosing pneumonia in X-ray images by more than 10%, a holdout blind test set found, with the AI more scalable. The accuracy exceeded benchmarks set by the current literature for that public data set.

"It can find the examples that are likely to be poor quality and flag those examples to a radiologist for further in depth review — serving to triage the cases that will be likely difficult to assess," said Perugini.

You Must Be Logged In To Post A Comment