Use of a deep learning Lung Cancer Prediction Convolutional Neural Network (LCP-CNN) algorithm in patients with solitary pulmonary nodules provides equivalent diagnostic accuracy to the current clinical standards of positron emission tomography/computed tomography (PET/CT) and dynamic contrast-enhanced CT (DCE-CT). This was among study findings published in Chest.
Study investigators hypothesized that the LCP-CNN would provide a high diagnostic accuracy that was unimpeded by the presence of an intravascular contrast agent. They suspected that use of advanced analytics of the baseline CT in which the nodule was detected could offer a more efficient approach to the workup of solitary pulmonary nodules at a lower cost and with less patient inconvenience. To test this hypothesis, the investigators conducted a post-hoc analysis that compared the diagnostic accuracy of LCP-CNN to values found in the SPUtNIK trial (ClinicalTrials.gov identifier: NCT02013063), a prospective, multicenter trial comparing the diagnostic accuracy of DCE-CT with PET/CT. The trial involved patients from multiple health care facilities in the United Kingdom.
Study inclusion criteria were a soft tissue solitary indeterminate nodule of at least 8 mm and 30 mm or less on axial plane measured on lung window via conventional CT scans, with no evidence that was strongly suggestive of malignancy. A total of 270 participants (51% male) underwent PET/CT and DCE/CT, with CT data available centrally to perform LCP-CNN analysis. The average patient age [SD] was 68.3 [8.8] years.
The LCP-CNN was designed and validated in an external cohort. LCP-CNN-generated risk scores were generated from the noncontrast and contrast-enhanced CT images from the DCE-CT. The gold standard used was histology or 2 years of follow-up.
Results of the study showed that the accuracy of the LCP-CNN on the noncontrast images (area under the receiver operator characteristic curve [AUROC], 0.83; 95% CI, 0.79-0.88) was superior to that of DCE-CT (AUROC, 0.76; 95% CI, 0.69-0.82; P =.03) and equivalent to that of PET/CT (AUROC, 0.86; 95% CI, 0.81-0.90; P =.35).
The presence of contrast was associated with a small decrease in diagnostic accuracy, with the AUROC declining from 0.83 (95% CI, 0.79-0.88) on the noncontrast images to 0.80-0.83 post-contrast (P <.05 for the difference between noncontrast and 240s post-contrast only).
Use of the LCP-CNN showed excellent interrater reliability (intraclass correlation coefficient [ICC], 0.994; 95% CI, 0.991-0.995) and interrater reliability (ICC, 0.999; 95% CI, 0.998-0.999), and was not affected by the presence of contrast.
Several limitations of the current study should be noted. The rate of malignancy in the SPUtNIK trial was 60%, which is relatively high. This rate is substantially higher than the rate reported in lung cancer CT screening programs. Additionally, 47 scans (12% of the entire cohort) were not available for analysis, which might have introduced selection bias.
“A LCP-CNN algorithm provides an AUROC equivalent to PET/CT in the diagnosis of solitary pulmonary nodules,” the study authors concluded. Although the LCP-CNN is impacted slightly by the presence of contrast, its accuracy remains high.
Disclosure: One of the study authors has declared affiliations with biotech, pharmaceutical, and/or device companies. Please see the original reference for a full list of the author’s disclosures.
Weir-McCall JR, Debruyn E, Harris S, et al; SPUtNIk Investigators. 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. Chest. Published online September 8, 2022. doi:10.1016/j.chest.2022.08.2227