Deep Image Prior with L1 Loss for Medical Image Denoising

This study introduces a Deep Image Prior (DIP)-based medical image denoising method that requires no training data, enhanced through the incorporation of an L1-norm fidelity term. The proposed model, referred to as L1-DIP, addresses the limitations of conventional DIP, which typically relies on an MSE loss and suffers from spectral bias and structural over-smoothing. By replacing the pixel-wise MSE loss with an L1 loss, the method improves robustness to outliers and better preserves anatomical discontinuities, making it particularly suitable for low-dose CT and high-resolution MRI reconstruction tasks. The model employs a U-Net equipped with skip connections, optimized per-image using a fixed random input. The L1 fidelity term guides the reconstruction process toward edge-aware solutions without requiring any external training data. Experiments were conducted on a thoracic CT scan and a coronal pelvic-lumbar MRI image, where additive Gaussian noise was synthetically added to simulate varying levels of corruption. Quantitative evaluations demonstrate that L1-DIP consistently outperforms DIP, DeepRED, and SURE-DIP across all noise levels, in terms of both peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). Visual comparisons confirm that the proposed model achieves clearer boundary definition and better preservation of anatomical structure. Convergence curves for PSNR and SSIM also reveal that L1-DIP converges faster than the DIP. These results highlight the effectiveness of integrating an L1-based fidelity term into the DIP framework, offering a training-free, high-fidelity solution for medical image denoising under complex noise conditions.