| dc.description.abstract | With the rapid development of COVID-19 into a global pandemic, Uganda is in dire need
of screening tools to aid risk strati cation for con rmatory testing, quarantining as well as
facilitate appropriate allocation of the limited health resources. Lately ultrasound has shown
relevance and dominance in respiratory screening due to its far-reaching availability, ease of
use and non-radiation competence. Lung ultrasonography (LUS) requires vivid knowledge of
identi cation of features i.e. A-lines which are a visual pattern which essentially represent a
healthy lung, B-lines, pleural e usion and lung consolidation which are prominent COVID-19
indicative artefacts. However, the accuracy and ease with which these artefacts can be spotted
is highly dependent on the expertise and experience of the sonographer. Therefore, there's
need for a uni ed pathway for the interpretation and classi cation of COVID-19 in LUS
images. In literature, deep learning has been shown to meet or exceed clinician performance
across most visual elds of medicine thus past work has been done in Localizing B-lines in
Lung Ultrasonography by Weakly-Supervised Deep Learning. This project involved detecting
and localizing B lines by employing modern deep learning strategies to support clinicians in
clinical assessments. However, this still involved interpretation which is subject to cognitive
bias and visual limitations. Thus, the proposed solution, machine learning- aided screening
of COVID-19 in LUS images which alleviates the observer variability among radiologists by
automation of classi cation of LUS images.
The machine learning-aided screening of COVID-19 LUS images project involves automatic
artefacts detection and interpretation which not only exceeds the limits of human vision
and cognitive bias but also robustly and accurately avails results on a web-based application
platform. This project has three contributions. First, we gathered a lung ultrasound dataset
consisting of 792 images (451 COVID-19 images and 341 HEALTHY images). This dataset
was assembled from various online sources, pre-processed speci cally for deep learning models
and split in a strati ed manner to obtain the train, validation and test set. Second, three
machine learning models i.e., VGG 16, VGG 19 and Resnet were successfully built, trained
and tested using the same binary class dataset. VGG 16 achieved an outstanding performance
with a classi fication accuracy of 98%, recall of 1, precision of 96%, F1 score of 97.82% and
ROC AUC of 99.9%. Lastly, the outperforming VGG 16 model was deployed on a decision
support web-based application.
1 | en_US |
