AI able to classify pure ground-glass opacities
Artificial neural networks (ANNs) can help radiologists classify ground-glass opacities (GGO) with improved accuracy, according to new findings published in Clinical Radiology.
“ANNs have proven useful in modelling complex relationships among predictor variables that are often obscure to clinical observation alone,” wrote J.A. Scott, MD, of the department of radiology at Massachusetts General Hospital in Boston, and colleagues. “Previous examples of ANN-based approaches to image interpretation have included lesion detection and evaluation, extraction of physiological parameters from radiography images, and differential diagnosis.”
The ANN was developed by separating 125 cases of pure GGOs at random, using 85 cases for training purposes and the other 40 as test cases. The ANN’s performance looking at 18FDG PET/CT studies was then compared to two subspecialty-trained radiologists, one with seven years of experience and another with 21 years of experience.
Overall, the ANN had an area under the ROC curve (AUC) of 0.981. The average predictions were 79.9% for malignant findings and 9.7% for benign findings. Its sensitivity was 100% (11/11 malignant lesions) and specificity was 93.1% (27/29 benign lesions). The radiologists, meanwhile, correctly identified 17 of the 29 benign lesions and found 23 lesions indeterminate.
“The ANN was able to correctly characterize a substantial percentage of pure GGOs as malignant, which the two expert readers could only classify as indeterminate. At a single point in time on a single chest CT … it is impossible for any radiologist, no matter how expert, to determine that a pure GGO is malignant. Some features of pure GGOs, however, can be used for their characterization as benign and were applied.”
Scott et al. also explained why ANNs perform so well in studies such as this one.
“A major advantage of ANN-based approaches to lesion diagnosis is their ability to capture often unsuspected and unrecognized interactions between disparate parameters that might be overlooked, if these parameters were considered in isolation,” they wrote. “For instance, although lesion size and morphology might each be associated with a particular likelihood of malignancy, a potential interaction between these two parameters might change the predictive value of either.”
Michael has more than 18 years of experience as a professional writer and editor. He has written at length about cardiology, radiology, artificial intelligence and other key healthcare topics.