Research Project 2020
Olin College of Engineering
Pictured above: a thumbnail set of x-ray scans exported by the program for review
Create a set of algorithms that uses computer vision to detect pneumonia in pediatric lung scans.
Two algorithms were implemented; one that uses linear regression to identify images and the other using deep learning.
For the linear algebra iteration of the model, I worked with another student at Olin College. I did a large portion of the MATLAB code, diagramming, and final paper. For the deep learning portion, I modified the ResNet-50 neural network implementation by Akhilesh Kumar.
Pneumonia is an inflammatory infection that targets the alveoli, small air sacs
in the lungs. This infection causes alveoli to fill with pus, causing a cough and
other influenza-like symptoms. This condition can be dangerous, as it decreases
the amount of oxygen that can reach a patient's bloodstream. Pneumonia can
be especially dangerous in children.
According to the World Health Organization, pneumonia accounts for 15% of
all deaths of children under 5 years old, killing more than 800,000 children in
2017. Considering the impact that is possible in this space, it is important
to have accurate and effective methods of detecting pneumonia for
A diagram showing the general process of the linear regression model
The implementation of the deep learning neural network significantly increased the accuracy of the process. After tweaking the linear regression model several times, we found a final accuracy of 79.75% this is promising, definitively better than a random guess. However, the deep learning method was far more accurate, delivering an accuracy of 97.44%
This staggering difference in accuracy, however, comes at a computational cost. With the linear regression model, the total time to "train" the image weights and check the dataset is around ten seconds. In the deep learning model, it takes more than two and a half hours to train the neural network to completion. That being said, the accuracy of the model crosses the ninety percent accuracy threshold within the first five minutes.
The confusion matrix of the deep learning method (97.4% accuracy)
Copyright © 2021 Avery Clowes - All Rights Reserved.