Comparison Of Manual And Synthetic Data Augmentation Based On Vision Transformer In Lung Disease Classification From Respiratory Sounds

Panjaitan, Ade Rizky (2026) Comparison Of Manual And Synthetic Data Augmentation Based On Vision Transformer In Lung Disease Classification From Respiratory Sounds. Undergraduate thesis, UPN Veteran Jawa Timur.

[img] Text (Cover)
22081010091-cover.pdf

Download (728kB)
[img] Text (Bab1)
22081010091-bab1.pdf

Download (121kB)
[img] Text (Bab2)
22081010091-bab2.pdf
Restricted to Repository staff only until 5 September 2028.

Download (1MB)
[img] Text (Bab3)
22081010091-bab3.pdf
Restricted to Repository staff only until 5 September 2028.

Download (2MB)
[img] Text (Bab4)
22081010091-bab4.pdf
Restricted to Repository staff only until 5 September 2028.

Download (1MB)
[img] Text (Bab5)
22081010091-bab5.pdf

Download (86kB)
[img] Text (Daftar Pustaka)
22081010091-daftarpustaka.pdf

Download (125kB)
[img] Text (Lampiran)
22081010091-lampiran.pdf
Restricted to Repository staff only until 2028.

Download (298kB)

Abstract

Class imbalance in respiratory sound datasets may cause classification models to favor majority classes and underperform on minority classes. This study investigates the effects of manual augmentation, MelGAN-based synthetic augmentation, and their combination on a Vision Transformer for classifying four respiratory sound classes in the ICBHI 2017 dataset: normal, crackle, wheeze, and both. Audio recordings were segmented, band-pass filtered, amplitude-normalized, resampled, standardized in duration, and converted into 224 × 224 log-Mel spectrograms. A ViT-B/16 model pretrained on ImageNet was trained through partial fine-tuning of the final six encoder blocks. Experiments covered original data, manual augmentation, synthetic MelGAN augmentation, and hybrid augmentation under patient-wise splits of 60:20:20, 70:15:15, and 80:10:10. The best model was selected using validation macro-F1. Synthetic MelGAN augmentation with a 70:15:15 split achieved the best performance, with 68.76% accuracy, 57.90% macro sensitivity, 85.50% macro specificity, and 59.99% macro-F1. Normal sounds were recognized most reliably, whereas crackle and both remained challenging. A supporting segment-level analysis yielded a macro-F1 of 58.74%. Overall, MelGAN augmentation provided the most balanced performance, while patient-wise splitting offered a stricter assessment of generalization to unseen patients.

Item Type: Thesis (Undergraduate)
Contributors:
ContributionContributorsNIDN/NIDKEmail
Thesis advisorAnggraeny, Fetty TriNIDN0711028201fettyanggraeny.if@upnjatim.ac.id
Thesis advisorPuspaningrum, Eva YuliaNIDN0005078908evapuspaningrum.if@upnjatim.ac.id
Subjects: Q Science > QA Mathematics > QA76.6 Computer Programming
Q Science > QA Mathematics > QA76.87 Neural computers
Divisions: Faculty of Computer Science > Departemen of Informatics
Depositing User: Ade Rizky Panjaitan
Date Deposited: 07 Sep 2026 02:14
Last Modified: 07 Sep 2026 03:17
URI: https://repository.upnjatim.ac.id/id/eprint/60005

Actions (login required)

View Item View Item