A hybrid frequency-spatial domain unsupervised denoising model for Gaussian-Poisson mixed noise in medical imaging

This paper proposes an unsupervised denoising model designed to address Gaussian-Poisson hybrid noise in CT, MRI, and X-ray images. Traditional deep image prior (DIP) methods suffer from slow convergence, spectral bias, and overfitting, limiting their clinical applicability. In this paper, by applying the Fourier transform, we incorporate frequency-domain priors extracted from the observed noisy image at the input stage. Instead of using both amplitude and phase, we rely solely on the amplitude spectrum, which captures the energy distribution of various frequency components while avoiding the instability associated with phase information. Meanwhile, we retain the spatial domain information to preserve the image's structural integrity, ensuring the effective capture of both low-frequency details and high-frequency anatomical features. This dual-domain strategy allows fine details to be captured early in training, thereby mitigating spectral bias, accelerating convergence, and improving the preservation of high-frequency anatomical structures. To further enhance diagnostic fidelity, we replace the conventional mean squared error (MSE) loss with an edge-aware L1 loss function that better preserves critical anatomical textures. Additionally, an entropy-based criterion tracks variations in image uncertainty over iterations to determine the optimal stopping point, effectively preventing overfitting without the need for external validation data. Experimental results demonstrate that our model achieves an average improvement of 10.7 % in PSNR and 17.9 % in SSIM compared to DIP, reaching peak performance in just 60 iterations, faster than the 1360 and 2990 iterations required by DIP and DIP-AITV, respectively. These findings highlight the efficiency and effectiveness of our method for medical image denoising.