Skip to Main Content
Principal Investigator
Lily Peng
Position Title
Program Manager
About this CDAS Project
NLST (Learn more about this study)
Project ID
Initial CDAS Request Approval
Apr 25, 2016
Developing automated algorithms for detection of lung pathology
Since 2013, the USPSTF recommends annual screening for lung cancer with low-dose computed tomography (LDCT) in adults aged 55 to 80 years who have a 30 pack-year smoking history and currently smoke or have quit within the past 15 years. However, access to health care and affordability issues still appear to be barriers to screening for lung cancer (Delmerico et al, 2014). Automating detection has the potential of increasing efficiency and reducing costs. Rapid advances in computer vision and large scale machine learning have made it possible to train computer algorithms to identify high-level concepts at an accuracy exceeding that of humans (Ioffe et al, 2015; Szegedy et al, 2015). Applying this approach to medical imaging, we have previously built an algorithm based on the Inception (aka GoogLeNet) models described in Szegedy et al, 2014 with promising results. For the current project, we aim to adapt these deep neural networks to analyze de-identified lung images (CT and X-ray). If successful, this work will help improve the detection of lung pathology, increase efficiency, and reduce costs associated with lung cancer screening in at risk patients.

Specific aim 1: Use deep learning techniques to classify de-identified lung radiologic studies (CTs, x-rays) according to the type of follow up that would be recommended by radiologists. Other input data may include serial data available on patients, the clinical and demographic metadata such as age, gender, family history, prior medical history and comorbidities, and smoking history. This would allow for risk stratification of patients being screened and guide the most appropriate surveillance and management.

Specific aim 2: Apply deep learning techniques to predict total and cause-specific morbidity and mortality from de-identified lung CTs and/or X-rays and the input data specified in aim 1.

Specific aim 3: Apply modern computer vision and deep learning methods, especially example-based super-resolution, to reconstruct a CT scan approximation from the scout image or from an x-ray. Determine if these CT scan approximations are sufficiently accurate to support aim 1 and aim 2. If successful this would significantly reduce radiation exposure and the associated cancer risks.


Daniel Tse, Google
Diego Ardila, Google
Atilla Kiraly, Google
Wenxing Ye, Google
Shravya Shetty, Google

Related Publications