The deep learning model to predict lung cancer risk within the next 3 years by once Low-dose computed tomography screening
Principal Investigator
Name
Lei Shi
Degrees
M.D.
Institution
Radiology Department of Zhejiang Cancer Hospital
Position Title
Chief Physician and Head of Radiology Department
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1265
Initial CDAS Request Approval
Jun 17, 2024
Title
The deep learning model to predict lung cancer risk within the next 3 years by once Low-dose computed tomography screening
Summary
Purpose:Low-dose computed tomography (LDCT) screening is a globally recognized effective method for reducing lung cancer mortality. Current diagnostic methods mainly based on nodule size and density, have shown limited accuracy. We hypothesized that deep learning algorithms could be employed to analyze LDCT images to predict individual risk of developing cancer within three years.
Methods:In this study, a deep learning algorithm was applied to analyze LDCT images from 19,823 participants in the screening program in Wenling, Zhejiang Province, China. Follow-up data over three years and histopathological examinations were used to identify 368 final cases of malignant tumors. Subsequently, a CT-based lung cancer risk prediction model was developed to predict the probability of cancer occurrence within three years for individuals. Given the low incidence of lung cancer in screening scenarios and the distinct differences in the size of benign and malignant nodules, 901 additional cases of benign and malignant nodules with matched sizes confirmed by pathology from Zhejiang Cancer Hospital were included to enhance the model's ability to distinguish between benign and malignant nodules. Independent testing was conducted on a population newly participating in screening the following year. Additionally, we aimed to perform independent external testing of the model on the National Lung Screening Trial (NLST) dataset.
Results:The lung cancer risk prediction model achieved an AUC of 0.93(95% CI, 0.90 to 0.96) for predicting lung cancer within three years in the cohort of individuals newly participating in screening the following year. The specificity was 95.8%, sensitivity was 69.1%, positive predictive value was 16.4%, and negative predictive value was 99.6%.
Conclusion:Our constructed lung cancer model can predict individual risk of developing cancer within three years based on a single LDCT scan, aiding in the optimization of lung cancer screening processes. This could potentially extend the interval for re-examination for some low-risk individuals, thereby reducing individual radiation doses and screening costs.
Methods:In this study, a deep learning algorithm was applied to analyze LDCT images from 19,823 participants in the screening program in Wenling, Zhejiang Province, China. Follow-up data over three years and histopathological examinations were used to identify 368 final cases of malignant tumors. Subsequently, a CT-based lung cancer risk prediction model was developed to predict the probability of cancer occurrence within three years for individuals. Given the low incidence of lung cancer in screening scenarios and the distinct differences in the size of benign and malignant nodules, 901 additional cases of benign and malignant nodules with matched sizes confirmed by pathology from Zhejiang Cancer Hospital were included to enhance the model's ability to distinguish between benign and malignant nodules. Independent testing was conducted on a population newly participating in screening the following year. Additionally, we aimed to perform independent external testing of the model on the National Lung Screening Trial (NLST) dataset.
Results:The lung cancer risk prediction model achieved an AUC of 0.93(95% CI, 0.90 to 0.96) for predicting lung cancer within three years in the cohort of individuals newly participating in screening the following year. The specificity was 95.8%, sensitivity was 69.1%, positive predictive value was 16.4%, and negative predictive value was 99.6%.
Conclusion:Our constructed lung cancer model can predict individual risk of developing cancer within three years based on a single LDCT scan, aiding in the optimization of lung cancer screening processes. This could potentially extend the interval for re-examination for some low-risk individuals, thereby reducing individual radiation doses and screening costs.
Aims
1.Construct a lung cancer prediction model under the LDCT screening scenario, achieving personalized and precise prediction of lung cancer risk on an individual basis, and validate it in actual screening scenarios.
2.Extend the LDCT follow-up time for some individuals based on the prediction results, thereby reducing radiation exposure and saving screening costs.
Collaborators
Chengting Lin,The Second Clinical Medical College, Zhejiang Chinese Medical University
Zhen Zhou,BEIJING DEEPWISE&LEAGUE OF PHD TECHNOLOGY CO.LTD
Weixiong Tan,BEIJING DEEPWISE&LEAGUE OF PHD TECHNOLOGY CO.LTD