By Will Boggs MD
NEW YORK (Reuters Health) - 11/7/2019
A new deep-learning framework based on CT images can help predict treatment failure and allow individualization of radiotherapy doses, researchers report.
"Our framework represents the first opportunity to use imaging with medical-record data to guide personalized radiation-dose recommendations," Dr. Mohamed E. Abazeed from the Cleveland Clinic, in Ohio, told Reuters Health by email. "We achieve this by recommending a dose of radiotherapy that is projected to minimize the risk of failure."
Deep-learning algorithms have shown promise for classifying patients on the basis of the type of cancer and genetic alterations, but there has been little progress in the use of the technique to predict tumor responses to individual anticancer therapies.
Dr. Abazeed and colleagues input pre-therapy lung CT images into Deep Profiler, a multitask deep neural network, and combined these data with clinical variables to derive iGray, an individualized radiation dose estimated to reduce failure probability at 24 months to below 5%.
Among 849 patients stratified by Deep Profiler, radiotherapy was more successful in the low-risk group (5.7% three-year cumulative incidence of local failure) than in the high-risk group (20.3% three-year cumulative incidence of local failure).
Deep Profiler predicted local failure regardless of the patient's cancer stage, the researchers report in The Lancet Digital Health, online July 1.
Models that included Deep Profiler and clinical variables predicted treatment failures with a C-index of 0.72, which was a significant improvement compared with classical radiomics or 3D volume.
iGray yielded optimal individual radiation doses ranging from 21.1 to 277 Gy and suggested dose reductions in 23.3% of patients.
As Dr. Abazeed envisions it, "the iGray dose recommendation will be reported along with a tunable bar that allows the physician to adjust the dose in order to achieve the most effective and safest dose based on the projected failure probability and prior knowledge about the potential for toxicity, respectively. By adding interpretability and reporting established clinical predictors, our goal is to minimize the 'black box' design of most interfaces that are built to translate machine-learning models into user-friendly medical-decision-support systems."
"Our framework could contribute, at least in part, to lowering healthcare disparities by providing guidance for treatment strategies in under-resourced medical facilities and populations," he said.
Dr. Raymond H. Mak from Dana-Farber Cancer Institute in Boston, who wrote an accompanying editorial, told Reuters Health by email, "The integration of clinical and medical-imaging data with radiation dose to predict likelihood of tumor response to radiation therapy has been a recognized clinical need for decades. The ability of a deep-learning approach to identify tumors at risk of recurrence with a performance that exceeded traditional measurements (tumor size/volume and radiomics) is impressive."
"We are on the cusp of a paradigm shift in cancer-outcomes prognostication and therapy-response prediction from traditional, human-derived measures of tumor size and volume on medical imaging to more automated deep-learning-driven methods," he said. "This study represents a watershed in that a hybrid approach is taken with the integration of human-derived features to build the AI model, but future approaches will likely continue the shift away from human-derived measures."
"I think we would need additional prospective testing before we can apply the deep-learning-based predictive model in the clinic directly," Dr. Mak cautioned. "In particular, I would caution that the model may not be easily generalizable to other cancer types and clinical situations."
"In this study," he added, "the model was trained on patients with relatively small lung cancers treated to very high doses of radiation therapy using the stereotactic body-radiation-therapy technique, which results in an unusually high likelihood of tumor control (about 90%). With this high-dose technique, it's unclear whether the model's predicted adjustments in dose would be clinically achievable. In particular, since these treatments are delivered with a goal of cure, any kind of dose de-escalation should be done on trial."
Siemens Medical Systems USA funded the study and employed several of the authors. Five authors are named inventors in a patent pending for the use of Deep Profiler and iGray to personalize radiotherapy doses.
Lancet Digital Health 2019.