A Power-Law noise scheduler for image diffusion models in low-dose CT denoising

Abstract

The growing need to reduce patient exposure to ionizing radiation in medical imaging has led to the widespread adoption of low-dose computed tomography protocols, guided by the principle of ALARA (As Low As Reasonably Achievable). However, lowering the radiation dose results in increased image noise and artifacts, which can significantly compromise diag nostic quality. Advanced deep learning techniques have demonstrated that neural networks can effectively learn mappings between LDCT and corresponding NDCT images, improving recon struction quality. Nevertheless, most existing approaches rely on direct mappings, which may have reached a performance plateau, as suggested by the diminishing improvements observed across recent models. In this context, diffusion models have emerged as promising alternatives for image restoration due to their generative capabilities. However, traditional formulations based on stochastic noise addition may not be well-suited when modeling complex noise dis tributions. Previous works have addressed this issue by redefining the forward process as a deterministic transformation between clean and noisy images, aligning more closely with ac tual real-world scenarios. Building upon this alternative formulation, we introduce a flexible power-law noise scheduler parameterized by an exponent γ, which controls the rate at which noise is introduced during the diffusion process. This design enables the exploration of different noise progression dynamics, and allows for the evaluation of Curriculum Learning hypotheses in the context of LDCT denoising. When γ > 1, the model follows a curriculum that gradually increases noise complexity, while γ < 1 corresponds to a reverse curriculum. This flexibility positions γ as a key component for guiding the learning process. Experimental results show that the proposed method outperforms standard reconstruction techniques, especially under aggressive noise schedules. Additionally, the best results were obtained with γ > 1, reinforcing the effectiveness of a curriculum-based degradation strategy in LDCT reconstruction tasks.