Radiation Image Denoising With DR-Specific Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing noise reduction methods in digital radiography apparatuses struggle to create appropriate rules for various subject structures, leading to inadequate noise reduction performance, and training neural networks for favorable noise reduction is difficult due to the unique noise characteristics of these systems.
Innovation Solution
An image processing apparatus that utilizes a learned model trained with radiation images containing artificially added noise with attenuated high-frequency components, employing a multilayer neural network to reduce noise in radiation images, including decorrelating and DC removing units to enhance noise reduction performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If rule-based noise reduction processing is used, then processing speed is maintained, but noise reduction performance is insufficient for various subject structures
Solution Approach 1:
The patent replaces the mechanical rule-based processing system with a neural network-based learning system. The noise reduction processing unit uses a trained neural network model to automatically learn and apply appropriate noise reduction for different subject structures, eliminating the need for manual rule creation and switching while achieving superior noise reduction performance across various imaging conditions.
2Reliability
If neural network training is performed with standard noise characteristics, then general noise reduction is achieved, but specific noise characteristics of digital radiography systems are not addressed
Solution Approach 1:
The patent performs preliminary action by pre-training the neural network with synthetic training data that specifically incorporates the noise characteristics of digital radiography systems (quantum noise and system noise). This pre-training preparation ensures that when the model is deployed, it immediately possesses the specialized knowledge needed to handle DR-specific noise patterns without requiring complex real-world data collection and annotation.
Solution Approach 2:
The patent introduces synthetic training data as an intermediary between the neural network and real-world DR images. This intermediary training data, generated with controlled noise characteristics matching DR systems, serves as a bridge that teaches the network the specific noise patterns it will encounter in practice, making the training process more efficient and effective.
3Ease of operation
If quantum noise and system noise are treated uniformly, then processing is simplified, but the different frequency characteristics of these noise types are not optimized
Solution Approach 1:
The patent applies local quality by enabling the neural network to automatically learn and apply different processing strategies for different noise types and their respective frequency characteristics. The network analyzes local image features and adapts its noise reduction approach accordingly, preserving high-frequency details while effectively reducing both quantum noise and system noise without uniform treatment.
Data Source
AI summary
An image processing apparatus is provided that includes: an obtaining unit is configured to obtain a first radiation image of an object to be examined; and a generating unit configured to, by inputting the first radiation image obtained by the obtaining unit into a learned model, generate a second radiation image in which noise is reduced compared to the first radiation image, wherein the learned model is obtained by training using training data that includes a radiation image obtained by adding noise with attenuated high-frequency components.