Computer Aided Diagnostic Tools for COVID-19 Detection via X-Ray Imaging
Presenter
Event Type
Workshop
Applications
Scientific Computing
Simulation
State of the Practice
Technology Challenge
W
TimeFriday, 13 November 20205:55pm - 6pm EDT
LocationTrack 10
DescriptionLightning talk: The goal of this study is to investigate lung geometry and density variations due to tissue and fluid alterations due to COVID-19. We approach this problem using two stages. We first have implemented neural networks (NN) as the key image processing method to solve a segmentation problem: detect both lungs. The Python-based algorithms explore public datasets as input to U-Net models, a convolutional NN architecture regulated by two main parameters, which influence performance: the number of downscaling (and subsequent upscaling) operations and the number of channels per feature map. Preliminary segmentation results show that the Dice score reached up to 0.946 using a 5-fold cross validation on two different datasets, (ieee8023 and NIH).
The second stage of our pipeline will involve using this segmentation to isolate the regions of interest as a preprocessing step for a second neural network that will be used to classify whether a patient has COVID-19 based on their Lung X-ray image. We intend to evaluate the ability of neural networks to classify lung images using a large set of 13000+ curated labeled lung X-ray images using a separated train and test set. We would like to compare the performances of various popular neural networks on this data set using the raw images as input as well as the segmented images as input in order to quantify the performance benefits of using segmentation as a preprocessing stage for the image analysis pipeline.
The second stage of our pipeline will involve using this segmentation to isolate the regions of interest as a preprocessing step for a second neural network that will be used to classify whether a patient has COVID-19 based on their Lung X-ray image. We intend to evaluate the ability of neural networks to classify lung images using a large set of 13000+ curated labeled lung X-ray images using a separated train and test set. We would like to compare the performances of various popular neural networks on this data set using the raw images as input as well as the segmented images as input in order to quantify the performance benefits of using segmentation as a preprocessing stage for the image analysis pipeline.