Automated measurement of malignancy risk of lung nodule detected by screening computed tomography

A. Ritchie, M. Tammemagi, C. Jacobs, W. Zhang, J. Mayo, H. Roberts, M. Gingras, S. Pasian, L. Stewart, S. Tsai, D. Manos, J. Seely, P. Burrowes, R. Bhatia, S. Atkar-Khattra, B. van Ginneken, M. Tsao and S. Lam

World Conference on Lung Cancer 2015.

BACKGROUND: We have previously reported a practical predictive tool that accurately estimates the probability of malignancy for lung nodules detected at baseline screening LDCT (New Engl J Med. 2013;369:908-17). Manual measurement of nodule dimensions and generation of malignancy risk scores is time consuming and subjected to intra- and inter-observer variability. OBJECTIVE: The goal of this study is to prepare a nodule malignancy risk prediction model based on automated computer generated nodule data and compare it to an established model based on radiologistsAC/a,!a,,C/ generated data. METHODS: Using the same published PanCan dataset (New Engl J Med. 2013;369:908-17) with the number of lung cancers updated, we prepared a logistic regression model predicting lung cancer using computer-generated imaging data from the CIRRUS Lung Screening software (Diagnostic Imaging Analysis Group, Nijmegen, The Netherlands). Ninety-one of the 2,537 baseline (first) scans were not available or could not be processed by CIRRUS. The remaining 2,446 scans were first annotated by the CIRRUS software. A human non-radiologist reader then accepted/rejected the annotated marks and manually searched the LDCT for nodules missed by CIRRUS or the study radiologist. New nodules found that were not recorded by the study radiologist were reviewed by a subspecialty trained chest radiologist with 14 years experience in lung cancer screening (JM). Nodule morphometric measurements (maximum and mean diameter, volume, mass, density) and total nodule count per scan irrespective of size were automatically generated by the CIRRUS software. The nodule type (nonsolid, part-solid, or solid), nodule description (lobulated, spiculated or well defined) and nodule location (upper versus middle or lower lobe) were manually entered. The variables were evaluated in models as untransformed and natural log transformed variables. Nonlinear relationships with lung cancer were also evaluated. Socio-demographic and clinical history predictors were not included in the model. RESULTS: Radiologists evaluation identified 8,570 pulmonary nodules of any size in 2063 individuals - 124 nodules in 119 individuals were diagnosed as cancer in follow-up. Based on CIRRUS software annotated marks that were accepted by a human reader, computer analysis identified 11,520 pulmonary nodules in 2174 individuals - 121 nodules in 115 individuals were diagnosed as cancer in follow-up. Thirty-six percent of the new nodules found by CIRRUS and/or second human reader were AC/aEURdegAY=4 mm (meanA,A+-SD, 5.9A,A+- 3.5 mm). Both the computer generated imaging data model (Model-CIRRUS) and the radiologist generated data model (Model-RAD) demonstrated excellent discrimination and calibration. Their predictive performances were also similar. Comparing Model-CAD to Model-RAD, the AUCs were 0.9537 versus 0.9541, the 90th percentile absolute errors were 0.0008 versus 0.0007, and the Brier scores were 0.0093 versus 0.0137. Mean nodule diameter is a better risk predictor than maximum nodule diameter, nodule density or mass. CONCLUSION: The predictive performances of computer and radiologist generated data models were similar. The model can be integrated to the CIRRUS Lung Screening software to automatically generate a nodule malignancy risk score to facilitate nodule management recommendation. ACKNOWLEDGEMENTS: Supported by the Terry Fox Research Institute, The Canadian Partnership Against Cancer and the BC Cancer Foundation.