Development and validation of a deep learning-based method for lung cancer risk prediction in low-dose chest CT images
Principal Investigator
Name
Haewon Kim
Degrees
M.D., Ph.D.
Institution
Keimyung University Dongsan Hospital
Position Title
Nuclear Medicine Associate Professor
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1097
Initial CDAS Request Approval
Jul 18, 2023
Title
Development and validation of a deep learning-based method for lung cancer risk prediction in low-dose chest CT images
Summary
The study aimed to develop a deep learning-based method for lung cancer risk prediction in low-dose computed tomography (LDCT) images. The method consists of a lung segmentation network and a 3D prediction network for lung cancer risk. The study population comprised individuals stratified based on their imaging results. The 2D densely connected U-Net was used for automatic lung segmentation, and 3D prediction networks were used to predict lung cancer risk in cropped LDCT images. The performance of 3D prediction networks such as 3D-CNN, MobileNet v2, SEResNet18, and EfficientNet-B0 were compared, with 3D-CNN achieving the highest AUC of 0.806 and sensitivity of 0.731. We expect that this approach could help in the selection of individuals at high risk for lung cancer for annual CT lung cancer screening.
Previously developed chest CT-based AI models focus on have already occurred lesions such as nodules, tumor, ILD, etc. However, the paradigm of the health care have changed and ultimate goals of the healthcare is the prevention. we developed an artificial intelligence algorithm that predicts the probability of lung cancer based on changes in the lung parenchyma that are currently invisible to the human eye. For developing our prediction model, we used the single-center dataset. therefore, we intend to perform external validation using the NLST dataset. through these processes, we expect to confirm the performance of the AI algorythm.
Previously developed chest CT-based AI models focus on have already occurred lesions such as nodules, tumor, ILD, etc. However, the paradigm of the health care have changed and ultimate goals of the healthcare is the prevention. we developed an artificial intelligence algorithm that predicts the probability of lung cancer based on changes in the lung parenchyma that are currently invisible to the human eye. For developing our prediction model, we used the single-center dataset. therefore, we intend to perform external validation using the NLST dataset. through these processes, we expect to confirm the performance of the AI algorythm.
Aims
- To analysis the NLST dataset and our dataset
- To compare the datasets and data features for fine-tunning the AI algorythms.
- To conduct the external validation of the developed deep learning-based model.
Collaborators
Haewon, Kim(Ph.D, MD, Neuclear medicine professor) Keimyung University Dongsan Hospital
Nowon, Kwon(MSN, RN, Clinical research team leader) Keimyung University Dongsan Hospital
Jungi, Lee(MS, AI developer) Keimyung University Dongsan Hospital
Sumin, Kim(BSN, RN, Clinical research team member) Keimyung University Dongsan Hospital
Hyehyun, Goo(BSN, RN, Clinical research team member) Keimyung University Dongsan Hospital