System and method for employing residual noise in deep learning denoising for x-ray imaging
Various methods and systems are provided for training a denoising system for a digital imaging system. The denoising system can be a deep learning denoising system formed as a blind or non-blind denoising system in which the training dataset provided to the denoising system includes a noisy image formed with simulated noise added to a clean digital image, and a reference image formed of the clean image having residual noise added thereto, where the residual noise is a fraction of the simulated noise used to form the noisy image. The use of the residual noise within the reference image of the training dataset teaches the DL network in the training process to remove less than all the noise during subsequent inferencing of digital images from the digital imaging system. By leaving selected amounts of noise in the digital images, the denoiser can be tuned to improve image attributes and texture.
1 . A method for training one of a blind denoiser or a non-blind denoiser for a digital imaging system, the method comprising the steps of:
a. generating training noise;
b. adding the training noise to a clean digital image to form a noisy image;
c. multiplying the training noise by a fraction to create a residual noise;
d. adding the residual noise to the clean digital image to form a reference image; and
e. providing the noisy image and the reference image to the denoiser as a training dataset,
wherein the training dataset including the residual noise in the reference image enables the denoiser to be tuned to remove more or less noise from a noisy image to adjust the strength of the denoising provided by the trained denoiser, and
wherein the digital imaging system is a digital medical imaging system.
2 . The method of claim 1 , wherein the training noise is image independent.
3 . The method of claim 1 , wherein the training noise is image dependent.
4 . The method of claim 1 , wherein the training noise can be extracted from actual images or simulated from a desired distribution.
5 . The method of claim 1 , wherein the step of generating training noise includes a transformation of the noise.
6 . The method of claim 5 , wherein the step of generating training noise includes a transformation of the noise to simulate the physical behavior of a system.
7 . The method of claim 6 , wherein the step of generating training noise includes a transformation of the noise to simulate the physical behavior of a component of a system.
8 . The method of claim 5 , in which the noise transformation is only applied to the noise added to form the reference image.
9 . The method of claim 5 , in which the noise transformation is only applied to the noise added to form the noisy image.
10 . The method of claim 5 , in which the noise transformation applied to form the noisy image is different than the noise transformation applied to form the reference image.
11 . The method of claim 1 , further comprising the step of forming a noise map as a part of the training dataset to achieve a non-blind denoiser.
12 . The method of claim 11 , wherein the training dataset includes a training mini batch created for a given input image, and wherein the mini batch set is formed in a method comprising the steps of:
a. generating N training noises;
b. generating N random fractions;
c. creating N residual noises by multiplying each training noise by a corresponding random fraction;
d. creating N reference images by adding each of the N residual noises to the input image;
e. creating N noisy images by adding the N training noise to the input image;
f. creating N noise maps;
g. employing the N noisy images, N reference images, N noise maps as a mini-batch set forming the training dataset; and
h. repeating steps a to g for each input image used for training.
13 . The method of claim 11 , wherein the training dataset includes a training mini batch set created for a given input image, and wherein the mini batch set is formed in a method comprising the steps of:
a. generating a training noise;
b. generating N random fractions;
c. creating N residual noises by multiplying the training noise by each N random fraction;
d. creating N reference images by adding each of the N residual noises to the input image;
e. creating N noisy images by adding the training noise to the input image;
f. creating N noise maps;
g. employing the N noisy images, N reference images, N noise maps as a mini-batch set forming the training dataset; and
h. repeating steps a. to g. for each input image used for training.
14 . The method of claim 1 , further comprising the step of loading the denoiser onto a digital imaging system after the denoiser is trained.
15 . The method of claim 1 , wherein the method is performed on the digital imaging system.
16 . The method of claim 1 , wherein the digital imaging system is a medical digital imaging system.
17 . A digital imaging system comprising:
a. a digital imaging source, and a digital imaging detector alignable with the digital imaging source;
b. an image processing system operably connected to the digital imaging source and digital imaging detector to generate digital image data, the image processing system including a processing unit for processing the digital image data from the detector;
c. non-transitory memory operably connected to the processing unit and storing instructions for operation of a denoiser and a denoiser training system,
d. a display operably connected to the image processing system for presenting information to a user, and
e. a user interface operably connected to the image processing system to enable user input to the image processing system,
wherein the denoiser is trained according the method of claim 1 .
18 . A digital imaging system comprising:
a. a digital imaging source, and a digital imaging detector alignable with the digital imaging source;
b. an image processing system operably connected to the digital imaging source and digital imaging detector to generate digital image data, the image processing system including a processing unit for processing the digital image data from the detector;
c. non-transitory memory operably connected to the processing unit and storing instructions for operation of a denoiser,
d. a display operably connected to the image processing system for presenting information to a user, and
e. a user interface operably connected to the image processing system to enable user input to the image processing system;
wherein the denoiser is trained using a denoiser training system and method configured to generate training noise, to add the training noise to a clean image to form a noisy image, to multiply the training noise by a fraction to create a residual noise, to add the residual noise to the clean image to form a reference image, and to provide the noisy image and the reference image to the denoiser as a training dataset,
wherein the training dataset including the residual noise in the reference image enables the denoiser to be tuned to remove more or less noise from a noisy image to adjust the strength of the denoising provided by the trained denoiser, and
wherein the digital imaging system is a digital medical imaging system.
19 . The digital imaging system of claim 18 , wherein the denoiser and denoiser training system is configured to operate on the digital imaging system or on a computing device remote from the digital imaging system.
20 . The digital imaging system of claim 18 , wherein the denoiser is loaded onto the digital imaging system after the denoiser is trained.