Researchers develop AI approach for high-risk clinical tumor volumes

por John R. Fischer, Senior Reporter | May 21, 2018
Rad Oncology Radiation Therapy
Researchers develop new AI approach
for contouring high-risk tumor volumes
in head and neck tumors
The risk of over- or under-dosing patients when establishing contouring plans may no longer be an issue with a technique that automates contouring for high-risk clinical target volumes.

Developed by researchers from The University of Texas MD Anderson Cancer Center, the computer-programmed approach utilizes AI and deep neural networks to replicate the patterns of physicians in treating specific types of tumors. If proven successful, the method offers great potential for treating head and neck cancers, for which contouring is a particularly sensitive task.

"For high-risk target volumes, a lot of times radiation oncologists use the existing gross tumor disease and apply a non-uniform distance margin based on the shape of the tumor and its adjacent tissues," Carlos Cardenas, a graduate research assistant and Ph.D. candidate at MD Anderson and the lead researcher behind the project, said in statement. "We started by investigating this first, using simple distance vectors."

Clinicians have shown much variability in how they contour clinical target volumes with a recent study conducted at Utrecht University showing differences in target volumes as large as eight times among colleagues.

Waiting is also a challenge, with an average of two to four hours needed for a radiation oncologist to determine clinical target volumes. Timing is further stalled at MD Anderson by the need for additional physicians to peer-review results to minimize the risk of missing diseases.

Researchers analyzed data from 52 patients with oropharyngeal cancer treated at MD Anderson between January 2006 and August 2010. Gross and clinical tumor volumes were previously contoured for radiotherapy.

Drawing from observations of the radiation oncology team at MD Anderson and how they target tumors, Cardenas began his research in 2015, eventually creating a deep learning algorithm with auto-encoders, a form of neural networks that can learn how to represent data sets, to comb through the data he collected.

Using the gross tumor volume and distance map information from surrounding anatomic structures as its input, the algorithm classified data to identify three-dimensional pixels known as voxels that are part of high-risk clinical target volumes.

Cardenas and his partners tested the approach on a group of cases that were not included in the training data, finding comparisons between their results and those of trained oncologists.

You Must Be Logged In To Post A Comment