Can Adjunctive AI Facilitate Earlier Lung Cancer Detection on Pre-Op CT Scans?

Adjunctive AI not only offers enhanced sensitivity for lung nodules, it may lead to earlier detection for patients who have had multiple preoperative computed tomography (CT) scans, according to emerging research findings.

For the retrospective study, recently reported in the European Journal of Radiology, researchers compared a CT-based artificial intelligence (AI) software (Veye Lung Nodules, version 3.9.2, Aidence) versus unassisted radiologist assessment of CT scans for 167 patients (mean age of 59) and a total of 475 resected nodules. All patients in the cohort had a lung metastasectomy, according to the study.

The study authors found that adjunctive AI facilitated a 92.4 percent sensitivity for preoperative detection of lung nodules on CT scans in contrast to 80.4 percent for unassisted radiologist interpretation. Specifically, the researchers noted the combination of radiologist assessment and adjunctive AI had a 97.3 percent sensitivity rate for metastatic nodules on preoperative CT in comparison to 89.7 percent for radiologist interpretation. The false positive rate with the AI software was 0.1 per CT scan, according to the study authors.

“In this study, we demonstrated that AI assistance has the capacity to significantly increase the radiologists’ sensitivity for the preoperative detection of lung nodules, at the cost of a very small number of false positives (FP),” wrote lead study author Giorgio Maria Masci, M.D., who is affiliated with the Radiology Department at Hospital Cochin in Paris, and colleagues.

In an analysis of 57 CT scans obtained prior to CT scans for initial reporting of lung nodules by radiologists, the study authors found that AI detected at least one nodule in 27 patients (47.4 percent). The AI software also detected metastatic nodules prior to radiologist detection in 21 of those patients (36.8 percent), according to the researchers.

“This advancement in the earlier diagnosis of metastatic status might impact patients’ management and has not been previously reported, to the best of our knowledge,” added Masci and colleagues.

In a multivariate analysis, the researchers found that nodules with vascular contact were commonly missed by radiologists (0.32 odds ratio (OR)) but also acknowledged that the AI software had low odds ratios for detection of cavitation (0.26) and nodules with a pleural base (0.10)

“ … The presence of cavitation or pleural contact was significantly associated with poor detection by the AI, suggesting that the deep learning software we used was not sufficiently trained to detect this type of nodule,” noted Masci and colleagues.

(Editor’s note: For related content, see "CT-Based AI Model May Enhance Prediction of Lung Cancer Recurrence," “Can Deep Learning Models Improve CT Differentiation of Small Solid Pulmonary Nodules?” and “FDA Clears CT-Based AI Software for Assessing Interstitial Lung Disease.”)

In regard to study limitations, the authors noted the lack of postoperative determination of missed metastases during surgical exploration precluded assessment of AI’s impact on the amount of resected metastases. They also noted the study’s emphasis on standard dose CT exams without evaluating the effect of reduced dosing on AI performance.