Model Deteksi COVID-19 dari Citra CT Scan Dada Menggunakan DenseNet-121
The primary diagnosis of COVID-19 is the RT-PCR test, but it was found that RT-PCR has the disadvantage of low sensitivity in the early phase of infection. Chest CT has the advantage of higher sensitivity in the early phase of infection compared to RT-PCR, so it can be used as a complement to the RT...
Saved in:
| Main Author: | |
|---|---|
| Format: | Theses and Dissertations NonPeerReviewed |
| Language: | English |
| Published: |
2024
|
| Subjects: | |
| Online Access: | https://repository.ugm.ac.id/285973/1/Model%20Deteksi%20COVID-19%20dari%20Citra%20CT%20Scan%20Dada%20Menggunakan%20DenseNet-121.pdf https://repository.ugm.ac.id/285973/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Universitas Gadjah Mada |
| Language: | English |
| Summary: | The primary diagnosis of COVID-19 is the RT-PCR test, but it was found that RT-PCR has the disadvantage of low sensitivity in the early phase of infection. Chest CT has the advantage of higher sensitivity in the early phase of infection compared to RT-PCR, so it can be used as a complement to the RT-PCR test to help reduce the spread of COVID-19 due to false negative results. To help medical personnel, Deep Learning can be used to automate the COVID-19 detection process via chest CT images.
In this research, a COVID-19 detection model was built by transfer learning of DenseNet-121. Several variations were done, that is without & with fine tuning, also variations on Learning Rate (LR) which was default LR (0.001) & LR obtained from Learning Rate Finder (0.0001). The model was trained using ReduceLROnPlateau & EarlyStopping callbacks. The dataset used was a dataset made of 3 classes (Normal, Pneumonia, & COVID-19) from COVIDx CT-2A which has gone through an undersampling process & various types of image augmentation. The model performance was then evaluated using various evaluation metrics namely accuracy, sensitivity, precision, & specificity.
The best results obtained were from the model with fine tuning & LR obtained from Learning Rate Finder. This model worked well, with an accuracy of 97.64%; precision of 96.49%; sensitivity of 96.43%; & specificity of 98.25%. |
|---|