| dc.description.abstract | Pneumonia, a lung disease often diagnosed using X-ray images, can be prone to
errors due to subjective interpretation. Machine learning methods offer a solution by
aiding clinicians in saving time and reducing diagnostic errors. CNN (Convolutional
Neural Network) architectures are neural network architectures specifically designed to
process spatial data such as images and videos. The CNN architectures used are: Basic
CNN, VGG16 and EffecientNet. In addition to that ViT is also used, ViT offers a
completely new approach to image processing using Transformer. Prior to training, the
dataset underwent preprocessing to augment its size and enhance model performance,
resulting in satisfactory average area under the curve (AUC) scores on the test set.
Following processing, the trained models were deployed on both an app and
website for easy accessibility by clinicians. ViT, a newer method primarily used in
research for pneumonia detection, demonstrated superior performance compared to
other methods, as evidenced by higher AUC scores. Specifically, ViT achieved an
accuracy of 95% on the training and validation sets, and 97.5% on the test set. Among
the four models examined, ViT demonstrates superior performance, succeeded by
EffecientNet, VGG16, and lastly the foundational CNN model. Specifically, despite
VGG16 being an antiquated model, its performance is commendable in terms of
stability and achieving high accuracy. VGG16 achieved an accuracy of 90% on the
training and validation sets, and 91.2% on the test set. The findings encompass four
models that hold promising potential for future applications in AI within the healthcare
domain.
In this study, deployment will be implemented on the model with the highest
accuracy, ViT, as well as the most stable model, VGG16. Despite ViT's strong
performance in training, VGG16 outperformed it when deployed to both apps and
webpages. The website in this study was found to be compatible only with VGG16,
with ViT unable to make predictions. This discrepancy may be attributed to differences
in the platforms of the two models. | en_US |
