A Hybrid Frequency-Spatial Domain Unsupervised Denoising Model for Gaussian-Poisson Mixed Noise in Medical Imaging

This paper introduces an unsupervised medical image denoising model designed to address the challenges of Gaussian-Poisson hybrid noise in CT, MRI, and X-ray images. Traditional deep image prior (DIP) methods, which use random noise as network input, suffer from slow convergence, while directly using observed noisy images often results in overfitting and poor denoising performance. To overcome these limitations, we fuse frequency and spatial domain information for multi-channel processing of the observed noisy images. A feature fusion module then integrates the benefits of both domains, enabling more accurate extraction of structural information, significantly improves denoising performance, and accelerates convergence. Furthermore, we introduce an entropy-based early stopping mechanism that dynamically monitors changes in entropy during training, automatically halting iterations once entropy decreases and stabilizes, thereby preventing overfitting. Additionally, the model employs an L1 loss function instead of the traditional MSE loss used in DIP to better preserve image edges and details. Experimental results demonstrate the proposed model's superiority, with an average increase of 10.7% in PSNR and 17.9% in SSIM compared to DIP. Remarkably, by the 60th iteration, the PSNR and SSIM values of the proposed method had already surpassed the peak values achieved by DIP at the 1,360th iteration and by DIP variants at the 2,600th iteration. The proposed model offers an efficient, robust, and innovative solution for medical image denoising tasks.