Contouring isn’t just for Kim Kardashian’s cheekbones.
Doctors use a process of the same name to determine high-precision radiation therapy treatments. Once the size and shape of tumors are identified in medical images, contouring helps them design target volumes for radiation therapy.
Researchers from the University of Texas MD Anderson Cancer Center, in Houston, are using deep learning to improve this time-consuming, labor-intensive process.
To contour, radiation oncologists review medical images of a patient, design specific target volumes of tumors and surrounding tissues, and decide how much radiation should be used.
As a result, there’s significant room for human error. If the tumor is in a vulnerable region, and a doctor administers too much radiotherapy, normal tissue could be damaged. If too little is administered, then the tumor may continue to grow.
AI Approach Aims to Narrow Variability
In its current state, contouring can also vary widely from doctor to doctor. A recent study from Utrecht University in the Netherlands found that some doctors suggested target volumes eight times larger than those decided upon by colleagues.
The MD Anderson researchers have developed a deep learning process to take on the tasks of contouring. Their AI-based approach would limit physician variability and prevent bias, decreasing the chance for missing a tumor or overtreating normal tissues.
The researchers used NVIDIA Tesla GPUs on the Maverick supercomputer at the Texas Advanced Computing Center (TACC) and the cuDNN-accelerated TensorFlow deep learning library to analyze data from 52 oropharyngeal cancer patients.
“If we were to do it on our local GPU, it would have taken two months,” said Carlos Cardenas, Ph.D., a researcher working with principal investigator Laurence Court, both at MD Anderson, in an interview with TACC. “But we were able to parallelize the process and do the optimization on each patient by sending those paths to TACC, and that’s where we found a lot of advantages by using the TACC system.”
The deep learning algorithm used neural networks to identify and recreate physician contouring patterns.
The process could have a significant impact on cancer treatments in general by making research and treatment more efficient. For example, researchers running clinical trials could use it to reduce uncertainties when delineating healthy tissues from cancerous ones in radiotherapy, which would in turn result in better clinical data.
Moreover, automating time-consuming processes like contouring and radiation treatment planning could greatly benefit patient care in clinics with low resources, many of which are found in low- and middle-income countries.
The researchers’ next steps are to translate this process into clinical use, ultimately providing physicians with better information that can lead to improved patient treatments and outcomes.
To that end, they’re developing a radiation planning assistant – a fully automated radiation therapy planning tool, which they hope to launch to partner cancer centers in low-income countries by early next year.