ParticipantsHyunho Park, Seoul, Korea, Republic Of (Presenter) Employee, VUNO Inc
Soo-Youn Ham, MD,PhD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Hwa-Young Kim, MD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Hyon Joo Kwag, MD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Seungho Lee, Seoul, Korea, Republic Of (Abstract Co-Author) Employee, VUNO Inc
Gwangbeen Park, Seoul, Korea, Republic Of (Abstract Co-Author) Employee, VUNO Inc
Sangkeun Kim, Seoul, Korea, Republic Of (Abstract Co-Author) Employee, VUNO Inc
Minsuk Park, Seoul, Korea, Republic Of (Abstract Co-Author) Employee, VUNO Inc
Jin-Kyeong Sung, MD, Seoul, Korea, Republic Of (Abstract Co-Author) Employee, VUNO Inc
Kyu-Hwan Jung, PhD, Seoul, Korea, Republic Of (Abstract Co-Author) Employee, VUNO Inc
hyunho.park@vuno.co
PURPOSETo evaluate the clinical value of a deep learning-based computer-aided detection (DLCAD) model that can reduce false negative reports on screening chest CTs that were considered normal.
METHOD AND MATERIALSA DLCAD consisting of a 2.5D CNN for candidate detection and a 3D CNN for false positive reduction was trained with a public LIDC-IDRI dataset. Preliminary validation performance for the same dataset was 90.7% sensitivity under one false-positive per scan threshold. Ten thousand low dose chest CT cases that were reported normal were collected from a single-center screening cohort from the year 2011 to 2015. 'Normal' was defined as containing no malignant or benign lesions. The deep learning-based CAD analyzed these cases reported as normal and detected nodule candidates. Four radiologists reviewed the results of CAD independently. When the candidate nodule was accepted, the type (solid, part-solid and ground-glass nodule [GGN]) and size of nodules were annotated.
RESULTSDLCAD analyzed 9952 cases (48 cases with inappropriate parameters, scan range or field of view were excluded) and detected 471 nodule candidates. Among them, 283 nodules from 269 patients were reported to be the true nodules by more than three radiologists. Excluding 67 nodules (with insufficient consensus), 216 nodules were categorized to be the same diameter range and nodule type by more than three radiologists. Among 216 nodules, 151 (69.9%) nodules were solid, three (1.4%) were part-solid, and 62 (28.7%) were GGN. Among 151 solid nodules, 10 (6.6%) nodules were larger than or equal to 6mm (eight [5.3%] 6 to 8mm, two [1.3%] 8 to 15mm) and 141 (93.4%) were smaller than 6mm. All three part-solid nodules were smaller than 6mm. All 62 GGN were smaller than 20mm. According to the Lung-RADS, two solid nodules were category 4A, eight solid nodules were category 3, and the remaining 206 nodules were category 2.
CONCLUSIONThe deep learning-based CAD has detected 2.7% (269/9952) false negative cases with neglected nodules. 4.6% (10/216) nodules were higher than Lung-RADS category 3, which require follow up scans.
CLINICAL RELEVANCE/APPLICATIONThe deep learning-based CAD will perform an ancillary role as a safeguard and a competent second reader by reducing false negative rates.