NoiseLens Evaluating Denoising for Diagnostic Information Preservation in Breast Ultrasound Imaging

Noise is an inherent challenge in ultrasound imaging and can affect the visibility of structures used for diagnosis. This is particularly important for breast cancer imaging in low-resource settings such as Nigeria, where breast cancer accounted for 25.3\% of reported cancer cases as of 2022. In response to this, healthcare spaces are increasingly exploring machine learning (ML) to support medical-image interpretation, clinical decision-making, and early detection. However, when ultrasound images are noisy, restoration may be required before they are passed to downstream ML models. This introduces an important concern for the reliability and clinical safety of the pipeline. While removing noise may improve image quality, the restoration process may also alter information used by downstream models. To investigate this, we compared three restoration approaches representing different modelling paradigms: DnCNN , a CNN-based architecture; Restormer, a Transformer architecture; and DMID , a diffusion method. The evaluation was done using the BUSI breast ultrasound dataset with the images synthetically degraded with Gaussian, speckle, and salt-and-pepper noise at four severity levels (0.05, 0.10, 0.20, and 0.40). The restored images were then evaluated using PSNR and SSIM for similarity to clean images, GLCM contrast for texture changes, and an EfficientNet-B0 classifier for downstream performance. The results showed that image similarity and preservation of information useful to a downstream model are not necessarily equivalent. For speckle noise, increasing degradation from 5\% to 40\% reduced classification accuracy from 90.3\% to 18.7\% without restoration. DnCNN largely prevented this decline, maintaining 97.4\% accuracy at both 5\% and 10\% degradation, 94.8\% at 20\%, and 90.3\% even at 40\% degradation. In contrast, Restormer followed the degradation trend, falling from 91.0\% to 19.4\%, while accuracy with DMID remained near 18--23\% across severity levels. This difference was also evident for salt-and-pepper noise, where DnCNN maintained 92.3--98.1\% accuracy across 5--40\% degradation, despite visible residual noise on the image after denoising. These results show that the effect of increasing image degradation on downstream performance was not determined by noise severity alone, but also by how the restoration method altered the image. Across the evaluated conditions, PSNR and SSIM did not consistently explain these differences, while greater GLCM contrast disruption generally coincided with lower classification accuracy. These findings suggest that conventional restoration metrics are not sufficient on their own to evaluate restoration as a preprocessing step for diagnostic ML. PSNR and SSIM describe similarity to a clean reference, while GLCM captures changes in texture, but neither directly indicates whether information useful to downstream analysis has been retained. Evaluating restoration alongside downstream model performance therefore provides a more complete assessment of what information remains available after restoration and highlights the need for more rigorous approaches to evaluating clinical information preservation in medical-image restoration.