Random Pixel Loss: Better Than Naive MSE for Medical Image Denoising

Medical imaging plays a crucial role in disease diagnosis, health monitoring, and treatment planning. The reliability of these processes depends heavily on the quality of medical images, which are often corrupted by noise introduced through human and machine errors. Removing this noise is essential, as it can otherwise lead to misinterpretation and misdiagnosis. Traditional denoising algorithms, while effective to an extent, often over smooth images, resulting in the loss of important structural details. In contrast, deep learning-based denoising methods have demonstrated superior performance, typically using Mean Squared Error (MSE) as their loss function. In this work, we introduce a novel loss function: Random Pixel Loss (RPL), as an enhancement to the MSE loss function. RPL places greater emphasis on the MSE of randomly selected pixels during loss estimation. This strategy improves the performance of the model. For evaluation, we implemented the UNet model as a backbone architecture. We validated RPL across multiple imaging modalities, including CT, MRI, OCT, and FUNDUS, and evaluated it against various types of noise such as Gaussian, Additive White Gaussian Noise (AWGN), and Speckle noise. The results demonstrate that RPL improves robustness and accuracy. Moreover, this approach can be integrated into existing state-of-the-art models to further boost their performance.