Electronic Theses and Dissertation

Analisis suara paru merupakan metode diagnostik non-invasif yang berperan penting dalam mendeteksi berbagai penyakit pernapasan. namun, kualitas rekaman suara paru sering mengalami penurunan akibat kontaminasi noise dari lingkungan maupun artefak fisiologis sehingga dapat mengurangi kualitas informasi diagnostik yang terkandung dalam sinyal. penelitian ini bertujuan mengimplementasikan metode denoising suara paru menggunakan kombinasi short-time fourier transform (stft) dan vision transformer (vit) untuk meningkatkan kualitas sinyal suara paru. tahapan penelitian meliputi prapemrosesan data berupa resampling, segmentasi, normalisasi amplitudo, serta penambahan synthetic noise berupa white noise, pink noise, dan heartbeat noise. selanjutnya, sinyal ditransformasikan ke domain waktu-frekuensi menggunakan stft sehingga menghasilkan spektrogram yang digunakan sebagai masukan model vision transformer. hasil prediksi kemudian direkonstruksi kembali ke domain waktu menggunakan inverse short-time fourier transform (istft). penelitian menggunakan icbhi 2017 respiratory sound database sebagai dataset utama, sedangkan evaluasi dilakukan menggunakan metrik signal-to-noise ratio (snr), mean square error (mse), dan short-time objective intelligibility (stoi) serta dibandingkan dengan metode wavelet. hasil penelitian menunjukkan bahwa metode stft-vit memberikan performa yang lebih baik dibandingkan metode wavelet. peningkatan kinerja terbaik diperoleh pada data pengujian pasien id icbhi 103, dengan perolehan snr sebesar 12,30 db, mse sebesar 0,004825, dan skor stoi sebesar 0,9859 dibandingkan kondisi sebelum denoising. analisis waveform, power spectral density (psd), dan spektrogram juga menunjukkan bahwa stft-vit lebih mampu mempertahankan karakteristik spektral suara paru sekaligus mereduksi energi noise. hasil penelitian menunjukkan bahwa kombinasi stft dan vision transformer merupakan pendekatan yang efektif sebagai metode denoising suara paru dan berpotensi diterapkan sebagai tahap prapemrosesan pada sistem analisis penyakit pernapasan berbasis kecerdasan buatan

Lung sound analysis is a non-invasive diagnostic approach that plays an important role in detecting various respiratory diseases. However, the quality of lung sound recordings is often degraded by environmental noise and physiological artifacts, reducing the diagnostic information contained in the signals. This study proposes a lung sound denoising method based on the combination of Short-Time Fourier Transform (STFT) and Vision Transformer (ViT) to improve lung sound quality. The proposed framework consists of data preprocessing, including resampling, segmentation, amplitude normalization, and synthetic noise injection using white noise, pink noise, and heartbeat noise. The audio signals are then transformed into the time-frequency domain using STFT to generate spectrograms, which are used as inputs to the Vision Transformer model. The predicted outputs are reconstructed into time-domain signals using the Inverse Short-Time Fourier Transform (ISTFT). The proposed method was evaluated using the ICBHI 2017 Respiratory Sound Database and compared with the conventional Wavelet denoising method. Performance evaluation was conducted using Signal-to-Noise Ratio (SNR), Mean Square Error (MSE), and Short-Time Objective Intelligibility (STOI). The experimental results demonstrate that the proposed STFT-ViT method consistently outperformed the Wavelet approach. The best performance was achieved on ICBHI patient ID 103, with an achievement of SNR about 12,30 dB, MSE 0.04825, and a STOI score of 0,9859 compared with the noisy input signal. Furthermore, waveform, Power Spectral Density (PSD), and spectrogram analyses indicate that STFT-ViT more effectively suppresses noise while preserving the spectral characteristics of lung sounds. These findings demonstrate that the integration of STFT and Vision Transformer provides an effective solution for lung sound denoising and has strong potential as a preprocessing stage in artificial intelligence-based respiratory disease analysis systems.