InnerEye is expected to cut radiotherapy planning time by up to 90%
AI solution expected to cut radiotherapy planning time by up to 90%
December 21, 2020
by John R. Fischer
, Senior Reporter
InnerEye, a new AI solution out of the University of Cambridge, has been found to be able to cut preparation time for radiotherapy by up to 90%.
It is expected to significantly reduce waiting times for beginning potentially lifesaving radiotherapy sessions. "Our national target is to ensure that patients receving curative radiotherapy must start within 28 days of the decision to offer treatment. We are looking at a range of system efficiencies to do much better than that," Dr. Rajesh JENA, group leader of the CRUK Cambridge RadNet Data Science Team, told HCB News.
“Helping people receive treatment faster is incredibly important and will not only improve recovery rates but will save clinicians precious time so they can focus on caring for patients,” said Health and Social Care Secretary Matt Hancock in a statement.
The solution was developed as part of an eight-year collaboration between the university, Microsoft Research and Addenbrooke’s Hospital. It is designed to save clinicians hours spent on manually marking up patient scans prior to radiotherapy, enabling shorter waiting times for treatment and saving providers money.
InnerEye ML models automate the image segmentation process in just a few minutes for precise targeting of tumors, while sparing healthy organs and tissues. It also is expected to reduce variability in findings brought on by the subjective nature of marking up patient scans.
When applied to an imaging data set of 519 pelvic and 242 head and neck CT scans among patients diagnosed with prostate or head and neck cancer, the models it produced were clinically accurate within the bounds of expert interobserver variability for 13 of 15 structures, and performed consistently well across previously unseen heterogeneous data sets. The correction time of autogenerated contours on 10 head and neck and 10 prostate scans were a mean of 4.98 minutes and 3.40 minutes per scan, respectively. This was shorter than the manual segmentation mean of 86.75 per scan by an expert reader of the head and neck scans, and the mean of 73.25 minutes per scan for a radiation oncologist. Autogenerated contours were produced in 93% of time required with manual contouring.
JENA says it the first time to his team's knowledge that an NHS Trust has implemented its own deep learning solution trained on its own data, which allows it to be used on patients and paves the way for more NHS Trusts to use open-source AI tools to help reduce cancer treatment times.
"First we are going through in-house regulatory approval to certify InnerEye as a medical device for use in our hospital," he said. "The work to do this is the same as the work to get CE Marking. At the end of the in-house approval process, we aim to apply for CE marking because it will really help other Hospitals deploy the tool."
He adds that the team has "retooled the InnerEye algorithms to look at Chest X-rays of cancer patients to pick up Covid when they come into hospital as an emergency, but that is another (emerging) piece of work."
Microsoft has made the InnerEye Deep Learning Toolkit freely available as open-source software.
The findings were published in JAMA Network Open.