Event Details
Detection of COVID-19 disease from X-ray images using convolutional capsule networks
Presenter: Donya Ashtiani Haghighi
Supervisor:
Date: Fri, August 12, 2022
Time: 10:00:00 - 11:00:00
Place: via Zoom - please see link below
ABSTRACT
Join Zoom Meeting
Register in advance for this meeting:
https://uvic.zoom.us/meeting/register/tZArdO6upzwsGdxG1rckNymDczCnn-sHkDJr
After registering, you will receive a confirmation email containing information about joining the meeting.
Meeting ID: 846 3769 7638
Note: Please log in to Zoom via SSO and your UVic Netlink ID.
ABSTRACT
Coronavirus (COVID-19) disease has spread abruptly all over the world since the end of 2019. The rapid and accurate diagnosis of COVID-19 is crucial for a better prognosis of this disease and breaking the chain of transition and flattening the epidemic curve. There are different types of COVID-19 diagnosis tests that sometimes have relatively low sensitivity. Computed tomography (CT) scans and X-ray images are other methods for the detection of this disease. However, one of the challenges of using these human-centered diagnosis methods is the overlap with other lung infections. Motivated by this challenge, different Deep Neural Network (DNN)-based diagnosis solutions have been developed, mainly based on Convolutional Neural Networks (CNNs), to accelerate the identification of COVID-19 cases. However, CNNs lose spatial information between image instances and require large datasets. In this seminar, an alternative framework based on Capsule Networks and Convolutional Neural Network is used which is able to handle small datasets. In addition, by investigating different parameters, the lowest loss of 0.0092, best accuracy of 0.9885, f1 score of 0.9883, the precision of 0.9859, recall of 0.9908, and Area Under the Curve (AUC) of 0.9948 is achieved when Plateau learning rate scheduler and margin loss function is used in capsule network. On the other hand, different dropout rates are used to decrease the overfitting, and the dropout rate of 0.1 shows better results. In the last part, by removing one capsule layer and having far less trainable parameters 146,752 in comparison to the main architecture, it still shows promising results.