According to the World Cancer Research Fund, lung cancer is the second most common form of cancer worldwide. Computed tomography (CT) is a three-dimensional (3D) imaging modality used often for cancer screening and diagnosis. The prime role of cancer screening is to determine the likelihood of developing cancer from the chest CT scans based on how indeterminate pulmonary nodules(IPNs) appear.
IPNs do not show distinct benign or malignant features and therefore, pose a diagnostic challenge to radiologists and clinicians. Previous studies have shown variable agreement among radiologists for IPN risk stratification and inconsistent adherence to established guidelines among pulmonologists. Several clinical risk estimation models have been developed for estimating IPN cancer risk. However, the results are not consistent.
As artificial intelligence (AI) technologies and machine learning algorithms continue to develop, they are becoming powerful tools in several fields including medicine. While CT scans can sensitively determine nodule characteristics such as border and size, AI can detect unique patterns predictive of cancer from very large datasets (which otherwise cannot be detected by the naked eye).
About the study
In the present retrospective multicenter study, researchers evaluated the performance of an AI-based lung cancer prediction convolutional neural network (LCP-CNN) computer-aided diagnosis (CAD) tool (developed by Optellum Ltd. of Oxford, England) for cancer risk estimation. They determined the inter-reader agreement for estimating cancer risks and management recommendations.
For the study, 5023 pulmonary CT scans with IPNs were obtained from seven different sources. Of which, 300 scans fulfilling the inclusion criteria were analyzed by a team of 12 readers comprising six pulmonologists and six radiologists. The IPNs appeared as rounded opacities (with 5 to 30mm diameter) and were surrounded by parenchyma. The readers allotted lung cancer prediction (LCP) scores to the IPNs for estimating cancer risks with and without the AI-based tool.
In addition, the team mentioned patient management recommendations which were: no action required, long-term (≥6 months) CT follow-up, short-term (six weeks to six months) CT follow-up, immediate additional imaging such as positron emission tomography/CT (PET/CT), non-surgical/ needle biopsy, or surgical excision
The CT scans were analyzed between June and July 2020 and data analysis was performed between August 2020 and December 2021. The study was published in the Radiology journal.
The mean age of the patients was 65 years and the majority (55%) of them were males. Fifty percent of the IPNs were estimated to be malignant (LCP 10) whereas the remaining half were benign (LCP 2). In comparison to benign pulmonary nodules, malignant nodules were associated with higher age (mean age, 68 years vs 62 years), women (54% vs 37%), partly solid density (12% vs 4%), higher nodule diameter (mean 12 mm vs 8.5 mm), and spiculation (53% vs 19%).
The AI-CAD tool improved readers’ estimation of nodule cancer risk [mean area under the curve (AUC) increased from 0.8 to 0.9], irrespective of reader specialty. At very low (5%) and high (65%) cancer risk thresholds, the integrated use of AI and CAD enhanced the diagnostic sensitivity from 94% to 98% and from 53% to 63 %, respectively.
Likewise, the corresponding diagnostic specificity increased from 37% to 42% and from 87% to 90%, respectively. The inter-reader agreement increased on AI-CAD use for both the 5% and 65% detection thresholds for cancer risk estimation. Similarly, the inter-reader agreement for management recommendations also improved with the AI-CAD tool.
The study showed that the AI tool enhanced the estimation of nodule cancer risk on chest CT when combined with clinical interpretations and improved inter-reader agreement for cancer risk stratification and patient management recommendations. It also indicates that the AI tool has promising potential to improve cancer diagnostics and patient management and boosts the incorporation of AI into pulmonary practice and radiology. The tool worked equally well on low-dose screening CT and diagnostic CT. However, future studies with real-world prospective clinical trials are required to accelerate its clinical translation.
Kim, R.Y., et al. (2022) Artificial Intelligence Tool for Assessment of Indeterminate Pulmonary Nodules Detected with CT. Radiology. doi.org/10.1148/radiol.212182.