Wavelet-enhanced autoencoder based image denoising with CNN fusion

In biomedical imaging, noise is a fundamental problem negatively impacting diagnostic quality, and effective noise reduction methods are critical for preserving structural details. This study proposes a hybrid noise reduction framework integrated with a CNN-based fusion network, combining a convolutional autoencoder, a specially designed hybrid wavelet filtering approach, and multi-level adaptive wavelet transformation. The autoencoder component learns clean image matches from noisy inputs, while the wavelet-based arm ensures the preservation of high-frequency structural and textural information through multi-level decomposition. The CNN-based fusion module adaptively combines these complementary representations to produce the final enhanced image. Experimental results demonstrate that the proposed method provides consistent superiority across different noise types and datasets. In the fetal ultrasound dataset, the highest performance was achieved at all levels under Gaussian and speckle noise; for speckle noise low-level PSNR values of 48.12 dB and SSIM of 0.99 were obtained. For brain tumor MR dataset, the method was effective not only under Gaussian and speckle noise but also under Rician noise; the PSNR value was obtained as 22.90 dB at the low level, 27.11 dB at the medium level, and 22.17 dB at the high level, with corresponding SSIM values of 0.90, 0.59, and 0.47. Furthermore, in non-referenced evaluations performed on the wrist trauma clinical X-ray dataset, mean NIQE = 3.12 and SSEQ = 0.26 values were obtained. These results confirm that the proposed approach offers balanced and generalizable denoising performance without excessive smoothing or the creation of artificial artifacts on real clinical images. These findings demonstrate that combining wavelet-based detail preservation with deep learning-based reconstruction delivers robust and reliable noise reduction performance across diverse biomedical imaging modalities and noise types.