Self-Fusion Simplex Noise-Based Diffusion Model for Self-Supervised Low-Dose Digital Radiography Denoising

Low-dose digital radiography (DR) is a commonly used imaging technique in clinical practice to minimize the harmful radiation for patients. At the same time, DR images are susceptible to noise and they often have degraded information contents, calling for effective image denoising techniques. The typical existing image denoising methods are supervised and they require clean samples at the training stage. Such requirement limits their clinical applicability since the clean samples are difficult to obtain in practice. To address this challenge, we propose a self-fusion diffusion model, denoted as S3-DDPM, with simplex noise in the frequency domain. Specifically, the proposed method first incorporates spectrum learning to capture rich image information in various frequency ranges without requirement of clean training samples. Second, it utilizes simplex noise instead of the typical Gaussian noise to enhance the diversity of the generated noise. Third, our method employs an alternating compensation strategy during each diffusion process, where a series of intermediate results are weighted and combined through self-fusion to derive a new denoised sample. The experimental results on clinical low-dose DR and CT images demonstrate the superiority of the proposed method over the state-of-the-art unsupervised methods, achieving favorable denoising performance.