Integrating AI into Radiology: Automated Detection of Pulmonary Emphysema
Principal Investigator
Name
Daniel Babalola
Degrees
B.S. in Economics
Institution
University of Pennsylvania
Position Title
Student
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1379
Initial CDAS Request Approval
Jan 10, 2025
Title
Integrating AI into Radiology: Automated Detection of Pulmonary Emphysema
Summary
This thesis aims to use deep learning to identify the presence of pulmonary emphysema in patients with a high risk of cancer. Emphysema is a chronic obstructive pulmonary disease (COPD) caused by chronic exposure to noxious gases According to the Global Burden of Disease Study in 2016, there were over 250 million cases of COPD worldwide; with 3.23 million deaths in 2019, it is the third leading cause of death globally. With smoking as the most common cause of this disease, it has a strong association with and can be an independent risk factor for lung cancer.
The model designed in this thesis will aim to serve a patient who has completed a low-dose CT scan (LDCT) for lung cancer screening and is found to have emphysema. Currently, the leading means of assessing emphysema in an LDCT is through direct visual analysis by an expert radiologist. Upon a diagnosis of emphysema, the doctor can conduct a spirometry test, a function test that measures airflow in the lungs, to assess the stage of the emphysema. Though radiologists are generally very effective at identifying the presence of emphysema in a LDCT, the implementation of a relatively accurate AI solution could provide the benefit of resource optimization in scenarios with limited access to trained radiologists or limited bandwidth of radiologists that could perform the assessment.
To accomplish this I will be taking a randomized sample of LDCT scans of patients with a history of heavy smoking, some of which will have emphysema, and training a deep learning model to accurately classify these patients. The model will apply a pre-trained convolutional neural network (CNN) architecture to extract feature maps of the images, with considerations being made for differences in image sizing, denoising, and normalization as necessary. With the feature maps extracted, the output will be integrated into a Shifted Window Transformer (Swin-T), a transformer model architecture designed for image classification that performs better than most other popular CNN models, which will classify the presence of emphysema.
The model designed in this thesis will aim to serve a patient who has completed a low-dose CT scan (LDCT) for lung cancer screening and is found to have emphysema. Currently, the leading means of assessing emphysema in an LDCT is through direct visual analysis by an expert radiologist. Upon a diagnosis of emphysema, the doctor can conduct a spirometry test, a function test that measures airflow in the lungs, to assess the stage of the emphysema. Though radiologists are generally very effective at identifying the presence of emphysema in a LDCT, the implementation of a relatively accurate AI solution could provide the benefit of resource optimization in scenarios with limited access to trained radiologists or limited bandwidth of radiologists that could perform the assessment.
To accomplish this I will be taking a randomized sample of LDCT scans of patients with a history of heavy smoking, some of which will have emphysema, and training a deep learning model to accurately classify these patients. The model will apply a pre-trained convolutional neural network (CNN) architecture to extract feature maps of the images, with considerations being made for differences in image sizing, denoising, and normalization as necessary. With the feature maps extracted, the output will be integrated into a Shifted Window Transformer (Swin-T), a transformer model architecture designed for image classification that performs better than most other popular CNN models, which will classify the presence of emphysema.
Aims
- Develop an image classification model that can identify emphysema in LDCT scans
- Test if this transformer model can achieve an accuracy of 95% or better.
- Test if this transformer model can be comparable to radiologists in accurate identification.
Collaborators
Dr. Eric Bressman MD, MSHP, Assistant Professor, Medicine (Hospital Medicine), Perelman School of Medicine