A comparative study to evaluate CT-based semantic and radiomic features in preoperative diagnosis of invasive pulmonary adenocarcinomas manifesting as subsolid nodules

Abstract

This study aims to predict the histological invasiveness of pulmonary adenocarcinoma spectrum manifesting with subsolid nodules ≦ 3 cm using the preoperative CT-based radiomic approach. A total of 186 patients with 203 SSNs confirmed with surgically pathologic proof were retrospectively reviewed from February 2016 to March 2020 for training cohort modeling. The validation cohort included 50 subjects with 57 SSNs confirmed with surgically pathologic proof from April 2020 to August 2020. CT-based radiomic features were extracted using an open-source software with 3D nodular volume segmentation manually. The association between CT-based conventional features/selected radiomic features and histological invasiveness of pulmonary adenocarcinoma status were analyzed. Diagnostic models were built using conventional CT features, selected radiomic CT features and experienced radiologists. In addition, we compared diagnostic performance between radiomic CT feature, conventional CT features and experienced radiologists. In the training cohort of 203 SSNs, there were 106 invasive lesions and 97 pre-invasive lesions. Logistic analysis identified that a selected radiomic feature named GLCM_Entropy_log10 was the predictor for histological invasiveness of pulmonary adenocarcinoma spectrum (OR: 38.081, 95% CI 2.735-530.309, p = 0.007). The sensitivity and specificity for predicting histological invasiveness of pulmonary adenocarcinoma spectrum using the cutoff value of CT-based radiomic parameter (GLCM_Entropy_log10) were 84.8% and 79.2% respectively (area under curve, 0.878). The diagnostic model of CT-based radiomic feature was compared to those of conventional CT feature (morphologic and quantitative) and three experienced radiologists. The diagnostic performance of radiomic feature was similar to those of the quantitative CT feature (nodular size and solid component, both lung and mediastinal window) in prediction invasive pulmonary adenocarcinoma (IPA). The AUC value of CT radiomic feature was higher than those of conventional CT morphologic feature and three experienced radiologists. The c-statistic of the training cohort model was 0.878 (95% CI 0.831-0.925) and 0.923 (0.854-0.991) in the validation cohort. Calibration was good in both cohorts. The diagnostic performance of CT-based radiomic feature is not inferior to solid component (lung and mediastinal window) and nodular size for predicting invasiveness. CT-based radiomic feature and nomogram could help to differentiate IPA lesions from preinvasive lesions in the both independent training and validation cohorts. The nomogram may help clinicians with decision making in the management of subsolid nodules.

Figure 1

Overall study design flowchart for the training and validation cohorts and diagnostic performance by each approach.

Figure 2

A typical example case of subsolid nodule with abnormal cystic-like airspace in LUL. A 65-year-old man had a 2.1 cm part-solid nodule with spiculated border in LUL. The (A) axial and (B) coronal CT images showed an abnormally dilated cystic-like airspace inside the lesion. The patient underwent video-thoracoscopic lobectomy of LUL. Further pathologic report demonstrated invasive pulmonary adenocarcinoma in LUL, Stage T1cN0M0. LUL left upper lobe.

Figure 3

A typical example case of subsolid nodule with an air bronchogram sign in LUL. A 61-year-old woman had a 1.4 cm part-solid nodule in LUL. The (A) axial and (B) coronal images showed an internal air bronchogram inside the lesion. The patient underwent video-thoracoscopic wedge resection of LUL. Further pathologic report demonstrated invasive pulmonary adenocarcinoma in LUL, Stage T1bN0M0. LUL left upper lobe.

Figure 4

Nomogram to predict the possibility of invasive pulmonary adenocarcinoma lesions based on GLCM-based feature (GLCM_Entropy_log10). To use the nomogram, an individual participant’s value is located on each variable axis, and a line is drawn upward to determine the number of points received for each variable value. The sum of these numbers is located on the total points axis to determine the possibility of invasive pulmonary adenocarcinoma lesions.

Figure 5

Calibration curves of the nomogram for predicting invasive pulmonary adenocarcinoma lesions from the training cohort. The Hosmer–Lemeshow test had a p value of 0.202 in the training cohort.

Figure 6

Calibration curves of the nomogram for predicting invasive pulmonary adenocarcinoma lesions from the validation cohort. The Hosmer–Lemeshow test had a p value of 0.917 in the validation cohort.