초록접수 현황
| 20F-034 | 구연 발표 |
Development of machine learning based algorithm for prediction of early lung adenocarcinoma by chest computed tomography
Kang Houn Lee₁, Ju Young Lee₁, Jin Sung Kim₁, Seong Yong Park₂
₁Department of Radiation Oncology, Yonsei Cancer Center, Severance Hospital, Yonsei University College of Medicine
₂Department of Thoracic and Cardiovascular Surgery, Yonsei University College of Medicine
Purpose : Discriminating subtypes of ground glass opacity nodule (GGN) between non-invasive / invasive pulmonary adenocarcinoma is clinically important for determining the extent of resection before surgery. We developed a machine learning based automatic segmentation and classification system with radiomic features to detect GGN and categorize its invasiveness
Methods : 190 Chest computed tomography (CT) sets and 205 GGNs were prepared from pathologically confirmed 205 cancer patients who had Lung resection surgery between 2018 and 2019. All GGNs were labeled and categorized by expert thoracic surgeon as atypical adenomatous hyperplasia (AAH), adenocarcinomas in situ (AIS), minimally invasive adenocarcinoma(MIA), and invasive adenocarcinoma (IA) based on the pathologic report. To facilitate pratical classification, GGNs were divided in two groups; noninvasive GGN (AAH, AIS and MIA) and invasive GGN (IA). We trained our networks on 163 ground glass opacity nodules with 6-fold cross validation and tested on 42 nodules. The Unet was used as main network and 'inceptionV3' as 'backbone' network for segmentation and 20 radiomic features are used for the IA classification.
Results : The mean age was 62.0 ± 9.6 years and male patients were 32%. Our proposed deep learning model for automatic segmentation scored 0.65 Dice-Coefficient score, 0.90 true positive rate, and false positive 1.5 (per case). The classification network using radiomics scored 0.90 AUC score, classifying between non-invasive/invasive GGN
Conclusion : Our preliminary machine learning based algorithm for automatic segmentation and classification showed acceptable outcome. Its usefulness has to be verified in the large data set.
Methods : 190 Chest computed tomography (CT) sets and 205 GGNs were prepared from pathologically confirmed 205 cancer patients who had Lung resection surgery between 2018 and 2019. All GGNs were labeled and categorized by expert thoracic surgeon as atypical adenomatous hyperplasia (AAH), adenocarcinomas in situ (AIS), minimally invasive adenocarcinoma(MIA), and invasive adenocarcinoma (IA) based on the pathologic report. To facilitate pratical classification, GGNs were divided in two groups; noninvasive GGN (AAH, AIS and MIA) and invasive GGN (IA). We trained our networks on 163 ground glass opacity nodules with 6-fold cross validation and tested on 42 nodules. The Unet was used as main network and 'inceptionV3' as 'backbone' network for segmentation and 20 radiomic features are used for the IA classification.
Results : The mean age was 62.0 ± 9.6 years and male patients were 32%. Our proposed deep learning model for automatic segmentation scored 0.65 Dice-Coefficient score, 0.90 true positive rate, and false positive 1.5 (per case). The classification network using radiomics scored 0.90 AUC score, classifying between non-invasive/invasive GGN
Conclusion : Our preliminary machine learning based algorithm for automatic segmentation and classification showed acceptable outcome. Its usefulness has to be verified in the large data set.
책임저자: Seong Yong Park
Department of Thoracic and Cardiovascular Surgery,Yonsei University College of Medicine
발표자: Kang Houn Lee, E-mail : dlrkdgns1324@gmail.com