AT AN AACR/IASLC JOINT CONFERENCE
SAN DIEGO (FRONTLINE MEDICAL NEWS) – A series of radiomics-derived imaging features may improve the diagnostic accuracy of low-dose CT lung cancer screening and help predict which nodules are at risk of becoming cancers.
“We are providing pretty compelling evidence that there is some utility in this science,” Matthew Schabath, Ph.D., said at a conference on lung cancer translational science sponsored by the American Association for Cancer Research and the International Association for the Study of Lung Cancer .
Radiomics is an emerging field that uses high-throughput extraction to identify hundreds of quantitative features from standard computed tomography (CT) images and mines that data to develop diagnostic, predictive, or prognostic models.
Radiologists first identify a region of interest (ROI) on the CT scan containing either the whole tumor or spatially explicit regions of the tumor called “habitats.” These ROIs are then segmented via computer software before being rendered in three dimensions. Quantitative features are extracted from the rendered volumes and entered into the models, along with other clinical and patient data.
“Right now our tool box is about 219, but by the end of the year we are hoping to have close to 1,000 radiomic features we can extract from a 3-D rendered nodule or tumor,” said Dr. Schabath, of the Moffitt Cancer Center in Tampa, Fla.
Although not without its own challenges, radiomics is a far cry from the current practice that relies on a single CT feature, nodule size, and clinical guidelines to evaluate and follow-up pulmonary nodules, none of which provides clinicians tools to accurately predict the risk or probability of lung cancer development.
CT images are typically thought of as pictures, but in radiomics, “the images are data. That’s really the underlying principle,” he said.
Led by Dr. Robert Gillies , often referred to as the father of radiomics, the researchers extracted and analyzed the 219 radiomic features from nodules in 196 lung cancer cases and in 392 controls who had a positive but benign nodule at the baseline scan and were matched for age, sex, smoking status, and race.
The post hoc, nested case-control study used images and data from the pivotal National Lung Screening Trial, which identified a 20% reduction in lung cancer mortality for low-dose CT screening compared with chest x-rays, but with a 96% false-positive rate, which also highlighted the challenges of LDCT as a screening tool.
Two classes of features were extracted from the images: semantic features, which are commonly used in radiology to describe ROIs, and agnostic features, which are mathematically extracted quantitative descriptors that capture lesion heterogeneity.
Univariable analyses were used to identify statistically significant features (threshold P value less than .05) and a backward elimination process (threshold P less than 0.1) performed to generate the final set of features, Dr. Schabath said.
Separate analyses were performed for predictive and diagnostic features.
In the risk prediction model, eight “highly informative features” were identified, Dr. Schabath said. Five were agnostic and three were semantic – circularity of the nodule, volume, and distance from or pleural attachment.
The receiver operating characteristic (ROC) area under the curve for the model was 0.92, with 75% sensitivity and 89% specificity. When the model included only patient demographics, it was no better than flipping a coin for predicting nodules at risk of becoming cancerous (ROC 0.58), he said.
Six highly informative features were identified in the agnostic model, which extracted features from the nodules found at the first and second follow-up interval, Dr. Schabath said. Three were agnostic and three semantic – longest diameter, volume, and distance from or pleural attachment.
The ROC for the diagnostic model was 0.89, with 74% sensitivity and 89% specificity.
When an additional analysis was performed using a nodule threshold of less than 15 mm to account for nodule growth over time and smaller nodule size at baseline in controls, the ROC and specificity held steady, but sensitivity dropped off to 59%, he said.
“I think we’re showing a rigorous [statistical] approach by identifying really unique, highly informative features,” Dr. Schabath concluded.
The overlap of volume and distance from or pleural attachment in both the diagnostic and predictive models suggests “there might be something very important about these two features,” he added.
Dr. Schabath stressed that the findings are preliminary and said additional analyses will be run before the results are ready for prime time. Long-term goals are to implement radiomic-based decision support tools and models into radiology reading rooms.
“In the future, we envision that all medical images will be converted to mineable data with the process of radiomics as part of standard of care,” Dr. Gillies said in an interview. “Such data have already shown promise to increase the precision and accuracy of diagnostic images, and hence, will increasingly be used in therapy decision support.”
Among the many challenges that first need to be resolved are that images are often captured with settings and filters that can be different even within a single institution. The inconsistency adds noise to the data that are extracted by computers.
“Hence, the most robust data we have today are generated by radiologists themselves, although this has its own challenges of being time-consuming with inter-reader variability,” Dr. Gillies noted.
Another major challenge is sharing of the image data. Right now, radiomics is practiced at only a few research hospitals and thus, building large cohort studies requires that the images be moved across site. In the future, the researchers anticipate that software can be deployed across sites to enable radiomic feature extraction, which would mean that only the extracted data will have to be shared, he said.