Training-method for a System for De-Noising Images

Simple SummaryContent extracted from patent full text and abstract with AI.

The patent describes a training method for a machine learning system designed to remove noise from images, particularly those acquired using techniques like magnetic resonance imaging (MRI). The method involves using multiple repetitions of complex-valued images, applying a phase correction to align them, generating a pixel-wise noise map, and then training the machine learning model using these processed images and the noise information with a special loss function (based on Stein's unbiased risk estimator) that does not require clean, noise-free images as ground truth. This approach improves the ability to de-noise images acquired with inherently low signal-to-noise ratios.

Use CasesContent extracted from patent full text and abstract with AI.

• Medical imaging, especially enhancing MRI images such as diffusion-weighted imaging (DWI), where noise is a significant issue.
• De-noising complex-valued images in other scientific imaging modalities (e.g., CT scans, PET scans, or other forms of advanced imaging where multiple repetitions and noise are common).
• Improving image quality in research settings where acquiring noise-free datasets is impractical or impossible.
• Integration into medical diagnostic devices and imaging software to provide clearer images for radiologists and clinicians.
• Development of software tools for automated image enhancement in clinical and research environments.

BenefitsContent extracted from patent full text and abstract with AI.

• Enables effective de-noising in images with inherently low signal-to-noise ratios, such as those from low-field MRI machines or diffusion-weighted imaging.
• Does not require noise-free 'ground truth' data sets for the training process, supporting fully unsupervised learning.
• Preserves important image details and structure while reducing noise, improving diagnostic quality and accuracy.
• Mitigates common problems like signal loss and motion-induced artifacts encountered in conventional averaging methods.
• Adaptable to a variety of image types and acquisition parameters, especially beneficial in medical imaging settings with variable protocols.
• Can be implemented in software and integrated into existing image reconstruction pipelines with relatively low overhead.

Technical Classifications (CPCs)

Inventors & Applicants

Patent Abstract

The invention describes a training-method for a system (12) with a machine learning model (20) for de-noising images, the training-method comprising the steps:- providing numerous image-datasets (D), wherein each image-dataset (D) comprises a plurality of complex valued image-repetitions (R),- performing a phase correction on the image-repetitions (R), wherein for each provided image-repetition (R) of an image-dataset (D) a phase corrected signal image (S) is calculated by amending the phase of the complex valued image-repetition (R) such that the phases of the image-repetitions (R) of the image-dataset (D) are consistent and such that the signal image (S) comprises signal contribution of the image-repetition (R),- calculating a noise map (M) for an image-dataset (D) based on the standard deviation between the signal images (S) of this image-dataset (D),- training the machine learning model (20) based on the signal images (S), the noise map (M) and a loss function (L) based on Stein's unbiased risk estimator.The invention further describes a related system, a filtering-method and a magnetic resonance imaging system.