Shin, Soogyeong (2024) Enhanced Disease Classification in Respiratory Sounds: A Transfer Learning Approach Utilizing ICBHI and Coswara Datasets. Master thesis, Voice Technology (VT).
|
PDF
MA-5661285-S-Shin.pdf Download (941kB) | Preview |
Abstract
The early detection of respiratory diseases such as asthma, chronic obstructive pulmonary disease, and pneumonia is vital for improving outcomes, a need highlighted by the COVID-19 pandemic. Traditional diagnostic methods, such as auscultation, depend on the experience of clinicians. To address this limitation, this study aims to increase the accuracy and efficiency of respiratory disease diagnosis using transfer learning through data augmentation. This research used publicly accessible respiratory sound datasets, ICBHI and Coswara, seeking to overcome the constraints of smaller datasets. A transfer learning employing a ResNet model was implemented to enhance the accuracy of respiratory disease diagnosis. The models, initially trained on ICBHI dataset, were fine-tuned with Coswara dataset, enabling them to detect abnormal respiratory sounds associated with specific disease. The results indicate that the model achieved accuracy peaks of 95.80% and 93% with the first and second fine-tuning strategies, respectively. These findings demonstrate the successful integration of two datasets through transfer learning. This study not only highlights highlights the importance of dataset augmentation to ensure robust model performance. Integrating multiple datasets through transfer learning can be an effective strategy not only for respiratory diseases but also for developing diagnostic tools for a variety of conditions with small datasets.
| Item Type: | Thesis (Master) |
|---|---|
| Name supervisor: | Verkhodanova, V. |
| Date Deposited: | 11 Jun 2024 07:24 |
| Last Modified: | 11 Jun 2024 07:24 |
| URI: | https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/466 |
Actions (login required)
 |
View Item |