Diagnostic Accuracy of a Convolutional Neural Network Assessment of Solitary Pulmonary Nodules Compared With PET With CT Imaging and Dynamic Contrast-Enhanced CT Imaging Using Unenhanced and Contrast-Enhanced CT Imaging

Abstract

Background: Solitary pulmonary nodules (SPNs) measuring 8 to 30 mm in diameter require further workup to determine the likelihood of malignancy.

Research question: What is the diagnostic performance of a lung cancer prediction convolutional neural network (LCP-CNN) in SPNs using unenhanced and contrast-enhanced CT imaging compared with the current clinical workup?

Study design and methods: This was a post hoc analysis of the Single Pulmonary Nodule Investigation: Accuracy and Cost-Effectiveness of Dynamic Contrast Enhanced Computed Tomography in the Characterisation of Solitary Pulmonary Nodules trial, a prospective multicenter study comparing the diagnostic accuracy of dynamic contrast-enhanced (DCE) CT imaging with PET imaging in SPNs. The LCP-CNN was designed and validated in an external cohort. LCP-CNN-generated risk scores were created from the noncontrast and contrast-enhanced CT scan images from the DCE CT imaging. The gold standard was histologic analysis or 2 years of follow-up. The area under the receiver operating characteristic curves (AUC) were calculated using LCP-CNN score, maximum standardized uptake value, and DCE CT scan maximum enhancement and were compared using the DeLong test.

Results: Two hundred seventy participants (mean ± SD age, 68.3 ± 8.8 years; 49% women) underwent PET with CT scan imaging and DCE CT imaging with CT scan data available centrally for LCP-CNN analysis. The accuracy of the LCP-CNN on the noncontrast images (AUC, 0.83; 95% CI, 0.79-0.88) was superior to that of DCE CT imaging (AUC, 0.76; 95% CI, 0.69-0.82; P = .03) and equal to that of PET with CT scan imaging (AUC, 0.86; 95% CI, 0.81-0.90; P = .35). The presence of contrast resulted in a small reduction in diagnostic accuracy, with the AUC falling from 0.83 (95% CI, 0.79-0.88) on the noncontrast images to 0.80 to 0.83 after contrast (P < .05 for 240 s after contrast only).

Interpretation: An LCP-CNN algorithm provides an AUC equivalent to PET with CT scan imaging in the diagnosis of solitary pulmonary nodules.

Trial registration: ClinicalTrials.gov Identifier; No.: NCT02013063.

Keywords: X-ray computed; diagnostic test accuracy; machine learning; positron emission tomography computed tomography; solitary pulmonary nodule; tomography.

Figure 1

Flow chart for the study. DCE = dynamic contrast-enhanced.

Figure 2

Graph showing comparison of the receiver operating characteristic curves for the LCP-CNN risk scores on the noncontrast and contrast-enhanced images. AUCs and between-group differences compared with the noncontrast LCP-CNN are as follows: noncontrast, 0.83 (95% CI, 0.79-0.88); 60 s after contrast, 0.81 (95% CI, 0.76-0.86; P = .07); 120 s after contrast, 0.81 (95% CI, 0.76-0.87; P = .05); 180 s after contrast, 0.83 (95% CI, 0.78-0.88; P = .56); and 240 s after contrast, 0.80 (95% CI, 0.75-0.86; P = .005). AUC = area under the receiver operating characteristic curve; DCE = dynamic contrast-enhanced; LCP-CNN = lung cancer prediction convolutional neural network.

Figure 3

Graphs showing comparison of the receiver operating characteristic curves for the LCP-CNN risk scores on the noncontrast scan, PET with CT scan imaging, and DCE CT imaging. AUCs and between-group differences compared with the noncontrast LCP-CNN as follows: noncontrast LCP-CNN AUC, 0.83 (95% CI, 0.79-0.88); DCE CT imaging maximum enhancement AUC, 0.76 (95% CI, 0.69-0.82; P = .03); PET with CT scan imaging SUV_max AUC, 0.86 (95% CI, 0.81-0.90; P = .35). AUC = area under the receiver operating characteristic curve; DCE = dynamic contrast-enhanced; LCP-CNN = lung cancer prediction convolutional neural network; SUV_max = maximum standardized uptake value.