Strip Diffusion Model for Low-Dose CT Image Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing low-dose CT image denoising methods face challenges due to the stringent clinical conditions for obtaining paired normal-dose and low-dose CT images, limited generality in handling different scanning protocols, and the loss of key details during interpolation, which affects image quality and subsequent diagnosis.
Innovation Solution
A zero-shot low-dose CT image denoising method based on a strip diffusion model, which involves building a strip diffusion model with a forward diffusion structure and a backward inference structure. The model processes normal-dose CT images to generate denoised low-dose CT images without requiring exact pairing of images, thus reducing data reliance and improving detail retention.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If supervised image post-processing methods are used for low-dose CT denoising, then denoising effectiveness is improved, but the requirement for paired normal-dose and low-dose CT images becomes extremely stringent, limiting practical application
Solution Approach 1:
The image is divided into multiple overlapping strips along the vertical direction. Each strip is processed independently by the diffusion model, which reduces the computational complexity and enables efficient parallel processing while maintaining overall image quality and detail preservation.
Solution Approach 2:
The method performs preliminary noise addition through forward diffusion process, then uses the trained diffusion model to predict and remove noise in the backward inference process. This two-stage approach allows the model to learn noise patterns effectively and apply denoising without requiring paired training data.
2Manufacturing precision
If direct prediction of diffusion model on entire image is performed, then denoising is achieved, but model training complexity significantly increases
Solution Approach 1:
The image is divided into multiple overlapping strips along the vertical direction. Each strip is processed independently by the diffusion model, which reduces the computational complexity and enables efficient parallel processing while maintaining overall image quality and detail preservation.
Solution Approach 2:
Instead of processing the entire image at once, the method processes partial regions (strips) independently with overlap. This partial action approach reduces computational burden and training complexity while the overlap ensures continuity and avoids edge effects between processed regions.
3Ease of manufacture
If interpolation methods are used for image processing, then image reconstruction is achieved, but key details are lost, bringing negative impact to subsequent diagnosis
Solution Approach 1:
The patent replaces traditional mechanical interpolation methods with a diffusion-based deep learning model. The diffusion model learns the underlying data distribution and generates realistic image details through probabilistic sampling, avoiding the blurring and artifact introduction characteristic of deterministic interpolation methods.
Solution Approach 2:
The method changes the fundamental approach from deterministic interpolation to probabilistic diffusion-based generation. By modeling the denoising process as a reverse diffusion trajectory, the system can generate high-frequency details that are consistent with the learned data distribution, preserving diagnostic information that interpolation would lose.
Data Source
AI summary
The present disclosure provides a zero-shot low-dose Computed Tomography (CT) image denoising method and apparatus based on strip diffusion model. This method includes building a strip diffusion model which uses high fidelity of the diffusion model to complete end-to-end denoising of low-dose CT images with different doses, thicknesses and devices. Particularly, in the training process, only the normal-dose CT images are required, greatly reducing the data reliance of the model, and model training across dose and thickness condition can be carried out with the data of only one scenario. In a sampling process, strip scanning strategy is used in combination with overlapped strip information and the input low-dose CT images to solve the maximum a posteriori problem, thereby sequentially generating denoising results. The present disclosure only uses simple convolution and attention architecture and carries out extensive experiments on the datasets involving different doses and thicknesses.