Deep learning-based differentiation of invasive adenocarcinomas from preinvasive or minimally invasive lesions among pulmonary subsolid nodules

Abstract

Objectives: To evaluate a deep learning-based model using model-generated segmentation masks to differentiate invasive pulmonary adenocarcinoma (IPA) from preinvasive lesions or minimally invasive adenocarcinoma (MIA) on CT, making comparisons with radiologist-derived measurements of solid portion size.

Methods: Four hundred eleven subsolid nodules (SSNs) (120 preinvasive lesions or MIAs and 291 IPAs) in 333 patients who underwent surgery between June 2010 and August 2016 were retrospectively included to develop the model (370 SSNs in 293 patients for training and 41 SSNs in 40 patients for tuning). Ninety SSNs of 2 cm or smaller (45 preinvasive lesions or MIAs and 45 IPAs) resected in 2018 formed a validation set. Six radiologists measured the solid portion of each nodule. Performances of the model and radiologists were assessed using receiver operating characteristics curve analysis.

Results: The deep learning model differentiated IPA from preinvasive lesions or MIA with areas under the curve (AUCs) of 0.914, 0.956, and 0.833 for the training, tuning, and validation sets, respectively. The mean AUC of the radiologists was 0.835 in the validation set, without significant differences between radiologists and the model (p = 0.97). The sensitivity, specificity, and accuracy of the model were 71% (32/45), 87% (39/45), and 79% (71/90), respectively, whereas the corresponding values of the radiologists were 75.2% (203/270), 76.7% (207/270), and 75.9% (410/540) with a 5-mm threshold for the solid portion size.

Conclusions: The performance of the model for differentiating IPA from preinvasive lesions or MIA was comparable to that of the radiologists' measurements of solid portion size.