<![CDATA[Malaria diagnosis accuracy depends on microscopic image quality, often compromised by noise. This study comprehensively evaluates classical denoising (morphological, median, bilateral filters) against deep learning architectures (DnCNN, Autoencoder, U-Net) for malaria parasite images. Using the Cell Images for Detecting Malaria dataset with synthetic Gaussian, salt-and-pepper, and mixed noise, experiments measured PSNR, SSIM, and processing time. Results indicate U-Net achieved superior performance (PSNR 36.69 dB, SSIM 0.9577), significantly outperforming Autoencoder (PSNR 26.12 dB) and classical methods (PSNR 23.14 dB). The baseline DnCNN architecture did not achieve competitive performance (PSNR 8.42 dB), indicating that domain-specific parameter tuning and data normalization adjustments are necessary for effective application to microscopic imaging. Autoencoder demonstrated the highest computational efficiency (1.64 ms per image), though the 10.57 dB PSNR gap relative to U-Net suggests that the quality trade-off may limit its suitability in accuracy-critical diagnostic scenarios. U-Net best preserved morphological details crucial for diagnosis and is recommended as the primary choice for malaria diagnostic systems prioritizing accuracy, while Autoencoder represents the most computationally efficient alternative for resource-constrained deployment. These findings support developing robust computer-aided diagnosis systems and contribute a comprehensive quantitative benchmark for denoising methods in malaria microscopy.]]>
