Skip to Main Content

An official website of the United States government

Principal Investigator
Jiang Yun
Peking University
Position Title
Director of Big Data and Machine Learning Innovation Center(MLIC), Associate Dean of EECS
About this CDAS Project
NLST (Learn more about this study)
Project ID
Initial CDAS Request Approval
Jul 13, 2015
Lung abnormalities detection via deep convolutional neural network
Classifying lung image with disease is a promising way for Medical diagnosis. Traditional methods require costly hand-crafted feature designing and the performance is far away from expectation. Recently, Convolutional Neural Network successfully attempts to model mammal visual cortex and improve the state-of-the-art of visual object recognition and object detection. However, the great success of Convolutional Neural Network heavily depends on a large amount of well-supervised data which is expensive to acquire. From our perspective, it is necessary to find a way to adapt the existing network to classify and detect lung abnormalities and take better advantage of the supervised information. Recent research (Li et al, 2014) has shown some promising techniques on lung pattern recognition based on Convolutional Neural Network. We plan to model the whole lung and learn to detect abnormalities, such as lung nodules.

Aim 1: Design or adapt a network framework based on existing models
How to build a network architecture based on the basic components is the key for promising performance. We plan to try some recently proposed techniques such as dropout, various pooling and nonlinear activation strategies to enhance the power of network. Based on the large amount of data provided by NLST, the parameters of the network will be learned well. Beyond that, we also want to design some data augmentation technique to make the network more robust.
Aim 2: Lung abnormalities detection
Based on the designed network framework, we will use some visualization technique to show the region of interest for the lung abnormalities and evaluate the results. The algorithm may also give some description of the abnormalities, such as the volume of lung nodules.