Photo Credit: Md Babul Hosen
Findings published in Academic Radiology suggest that deep learning (DL) based on CT imaging holds promise for distinguishing nasal polyps (NP) from inverted papilloma (IP) lesions. Study investigators used three DL models (3D ResNet, 3D Xception, HRNet) to identify NP from IP lesions in nearly 1,800 patients with nasal benign tumors. Patients were divided into training, internal test, and external test sets. The best-performing DL model was 3D Xception, with an area under the receiver operating characteristic curve of 0.999 in the training, 0.981 in the internal, and 0.933 in the external test sets, as well as sensitivity and specificity that were greater than those of four radiologists. This DL model also improved radiologist sensitivity from 0.845 to 0.884 and specificity from 0.670 to 0.840, and predictions from the model were associated with epithelial cell differentiation.
