
Contact the VUNO team anytime.
RSNA 2020
Purpose To evaluate the performance of deep-learning based algorithm (DLA) for auto-measurement of solid portion in surgically proven lung adenocarcinomas with ground glass opacity (GGO).
Method and Materials From January 2018 to December 2018, 432 surgically proven lung adenocarcinomas manifesting GGO portion at CT were included. Five radiologists independently measured the maximum axial diameter of solid portion in lesions on lung window. The deep-learning based algorithm automatically segmented and measured the axial, coronal, and sagittal maximum diameter of solid portion on lung window. The reader, software measurement, and invasive size of pathologic report were compared with each other.
Results 432 lesions consisted of 396 part-solid lesions and 36 pure GGO lesions, with the invasive size on pathology ranged from 6 to 65 mm (median, 18 mm). The agreement of measurement on axial plane between readers and DLA were very good with a correlation coefficient (ICC) of 0.877, while the ICC among 5 readers were 0.904. The range of mean difference was -4.02 – 1.41 mm (the range of lower and upper 95% limits of agreement [LOA], -18.12 – -10.48 and 10.08 – 14.16, respectively) when compared each readers with DLA and -5.42 – 1.02 mm (the range of lower and upper 95% LOA, -16.18 – -8.78 and 5.04 – 10.82, respectively) when compared readers each other. The agreement between DLA and invasive size on pathology was good with ICC of 0.758.
Conclusion Auto-measurement of solid portion in lung cancer with GGO showed good agreement with manual measurement and invasive size on pathology.
Clinical Relevance/Application The automated measurement of solid portion in pulmonary lesions can allow effective and reproducible assessment of lesion and subsequently facilitate automated pulmonary nodule management according to Lung-RADS or other guidelines.