Skip to Main Content

An official website of the United States government

Principal Investigator
Shipeng Xie
1.Nanjing University of Posts and Telecommunications 2.University of Pennsylvania
Position Title
Visitor Scholar
About this CDAS Project
NLST (Learn more about this study)
Project ID
Initial CDAS Request Approval
Aug 31, 2016
Classification of lung lesions using a convolutional neural network
In this study, a method identifying lung nodules automatically is proposed. The method employs a convolutional neural network (CNN) that labels each lung (left lung and right lung) as containing lung nodules or not. The analysis is performed in axial slices following NLST’s settings.

Each lung will be automatically roughly delineated by Jayaram K. Udupa’s method [1]. Then, the CNN will be used to identify whether the lung has nodules or not. To automatically determine presence of lung nodules, patches with a size of 48 x 48 pixels will be extracted from the axial slices. Patch size was chosen to ensure that the lung nodules would be contained within the patch. The CNN was designed with three convolutional layers each made of 32 filters, followed by a sub-sampling 2 x 2 max-pooling layer. Resulting feature maps will be connected to two fully-connected layers with 16 and 4 neurons, respectively, followed by a softmax output layer to produce probabilities for the given patch. Sub-sampling layers will be used to reduce the size of the resulting feature maps and to reduce the number of optimization parameters. In order to reduce the chance of over-fitting, the “dropout” strategy will be used. Batch normalization will be applied to accelerate the training process.

[1] Udupa, J. K., et al. (2014). "Body-wide hierarchical fuzzy modeling, recognition, and delineation of anatomy in medical images." Med Image Anal 18(5): 752-771.

Use a convolutional neural network to identify lung lesions.


Jayaram K. Udupa, Medical Image Process Group, University of pennsyvania