NUST Institutions Library Catalogue

Accurate analysis of respiratory sounds holds key importance in precise diagnosis of pulmonary
diseases. These sounds are sometimes noisy which require an effective diagnosing method for
noise reduction and correct analysis of lung sounds. Traditional denoising methods are limited to
spectral overlap with background noise. This study attempts for denoising biomedical signals
utilizing discrete wavelet convolutional neural network (DW-CNN), an adaptive filter consists of
multiresolution ability of discrete wavelet transform (DWT) and deep feature learning ability of
neural networks, which preserve the signal details. Encoder-decoder structure of DW-CNN with
inverse DWT accurately reconstructs the signal, outperforming traditional wavelet denoising. LS
signals under real noise conditions were denoised using DW-CNN, where DWT replaced
traditional pooling layers. Results were quantitatively compared with baseline traditional methods
using Signal to Noise Ratio (SNR) and Root Mean Square Error (RMSE) metrics. On average
there is an improvement of 9.61 dB in SNR and a reduction of 0.35489 in RMSE. Final results
highlight the better performance of implemented model over conventional methods, increasing the
scope of combined model in pattern recognition, clinical diagnostics and wearable devices.