ParticipantsMichelle L. Hershman, MD, Tucson, AZ (Presenter) Nothing to Disclose
Ryan J. Avery, MD, Tucson, AZ (Abstract Co-Author) Nothing to Disclose
Diana M. Palacio, MD, Tucson, AZ (Abstract Co-Author) Nothing to Disclose
Berndt P. Schmit, MD, Tucson, AZ (Abstract Co-Author) Nothing to Disclose
Matthias Wolff, MD, Erftstadt, Germany (Abstract Co-Author) Nothing to Disclose
Luca Bogoni, PhD, Malvern, PA (Abstract Co-Author) Employee, Siemens AG
mlhershman@gmail.com
PURPOSETo assess if nodule detection using CAD for lung cancer screening as an initial reader may correctly determine Lung-RADs categorization with suitable sensitivity and low false-positive rate.
METHOD AND MATERIALSA retrospective study of 86 low-dose CT lung screening exams were first assessed using proposed CAD marks by 3 independent readers of variable experience (novice-expert) and then reviewed for any additional nodules. For each finding, readers specified nodule type, diameter and confidence level (1= "not a nodule", 10= "definitively a nodule"). The standard of reference was determined by majority agreement and arbitrated by an expert chest radiologist. The average reader sensitivity, specificity and AUC were calculated with respect to nodule as well as LRAD cutoff scores of 2 and 3. CAD standalone sensitivity and CAD and average reader per case false-positive rates were computed. Standard errors and 95% confidence intervals were derived from 1000 bootstrap samples, which were used to derive p-value comparing CAD alone and CAD plus reader.
RESULTSCAD and/or readers identified 505 findings. True nodules (n=119) included findings >=4mm, confidence >=7.5 and were part-calcified, solid or subsolid. Smaller and calcified nodules and lymph nodes were excluded (n=195). Sensitivity of CAD alone was 86% (95% CI, 80-92%) compared to the 3 reader average which was 64% (95% CI 58-69%). False-positives per case for CAD alone was 1.952 (95% CI 1.506-2.446) while the reader average was 0.088 (95% CI 0.052-0.129). Reader average specificity was 96% (95% CI 95-98%) and AUC was 0.799 (95% CI 0.770-0.829) accounting for the 191 false-positives. AUC using LRAD >=2 cutoff demonstrated no significant difference (p=0.80) for CAD alone (0.847, 95% CI 0.795-0.892) compared to CAD plus reader average (0.844, 95% CI 0.796-0.886). Using LRAD >=3 cutoff, AUC also demonstrated no significant difference (p=0.47) for CAD alone (0.821, 95% CI 0.754-0.884) versus CAD plus reader average (0.839, 95% CI 0.778-0.898).
CONCLUSIONUsing a workflow with CAD as a first reader followed by radiologist review is feasible for determining Lung-RADs categorization and management.
CLINICAL RELEVANCE/APPLICATIONComputer-aided detection can serve as a first reader for detection of pulmonary nodules on lung cancer screening CT exams, producing improved sensitivity to the human reader with a very low false-positive rate and no significant changes in patient management.