2.5d window attention low-dose ct denoising method

CN121746543BActive Publication Date: 2026-06-09ZHEJIANG UNIV

Patent Information

Authority / Receiving Office: CN · China
Patent Type: Patents(China)
Current Assignee / Owner: ZHEJIANG UNIV
Filing Date: 2026-03-02
Publication Date: 2026-06-09

AI Technical Summary

Technical Problem

Low-dose CT imaging is characterized by high noise and obvious artifacts. Existing 2D methods struggle to utilize axial redundancy information, while 3D methods have high computational costs and are difficult to balance between noise suppression and detail preservation.

Method used

A 2.5D multi-channel input and window attention Transformer network (RestormerWIN) is adopted. Through a multi-scale encoder-decoder structure and a windowed attention module, it explicitly utilizes the information of adjacent slices and combines a gated depthwise convolutional feedforward module to achieve noise suppression and detail preservation.

Benefits of technology

Without significantly increasing computational overhead, it improves the discernibility of fine structures, enhances image quality, reduces noise and artifacts, and adapts to deployment requirements under different computing power conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

Figure CN121746543B_ABST

Patent Text Reader

Abstract

The application discloses a 2.5D window attention low-dose CT denoising method, and belongs to the field of medical image processing and artificial intelligence. The application firstly makes a training set containing paired low-dose CT slices and reference high-quality CT slices; then uniformly pretreats the low-dose CT slices; then constructs a 2.5D multi-channel input for a target low-dose CT slice, inputs the 2.5D multi-channel input into a RestormerWIN multi-scale encoding-decoding network for forward inference, and obtains a denoising output; calculates the error between the denoising output and the reference high-quality CT slice, optimizes the parameters of the network, and obtains a trained denoising model; for a low-dose CT image to be denoised, the low-dose CT image is pretreated, and then the trained denoising model is used to generate a denoised CT image. The application realizes effective suppression and detail preservation of LDCT noise and artifacts; and improves the deployment adaptability of the model under different computing power conditions.

Need to check novelty before this filing date? Find Prior Art