Systems and methods for noise reduction in medical images with deep neural networks
Methods and systems are provided for reducing noise in medical images with deep neural networks. In one embodiment, a method for training a neural network comprises transforming each of a plurality of initial image data sets not acquired by a medical imaging modality into a target image data set, wherein each target image data set is in a format specific to the medical imaging modality, corrupting each target image data set to generate a corrupted image data set, and training the neural network to map each corrupted image data set to the corresponding target image data set. In this way, the high-resolution of digital non-medical photographs or images can be leveraged for the enhancement or correction of medical images, and the trained neural network can be used to reduce noise and image artifacts in medical images acquired by the medical imaging modality.
1. A method of training a neural network, the method comprising:
transforming each of a plurality of initial image data sets not acquired by a medical imaging modality into a target image data set, wherein each target image data set is in a format specific to the medical imaging modality;
corrupting each target image data set to generate a corrupted image data set; and
training the neural network to map each corrupted image data set to the corresponding target image data set.
2. The method of claim 1 , wherein the medical imaging modality is magnetic resonance (MR), the initial image data sets are digital photographic data sets, and transforming each initial image data set into a target image data set comprises converting color or grayscale channels of each digital photographic data set into an MR image data set of complex values.
3. The method of claim 2 , further comprising adding background phase to the MR image data set of complex values.
4. The method of claim 1 , wherein the medical imaging modality is MR, the initial image data sets are digital photographic data sets, and transforming each initial image data set into a target image data set comprises converting color or grayscale channels of each digital photographic data set into an MR parametric map that models one or more of proton density, relaxation time, magnetic susceptibility, chemical shift, temperature, and diffusivity.
5. The method of claim 1 , wherein the medical imaging modality is ultrasound, the initial image data sets are digital photographic data sets, and transforming each initial image data set into a target image data set comprises converting color or grayscale channels of each digital photographic data set into an ultrasound parametric map that models one or more of acoustic impedance, velocity, and density.
6. The method of claim 1 , wherein the medical imaging modality is computed tomography (CT), the initial image data sets are digital photographic data sets, and transforming each initial image data set into a target image data set comprises converting color or grayscale channels of each digital photographic data set into a CT parametric map that models radiation absorption.
7. The method of claim 1 , wherein the imaging modality is MR, and corrupting each target image data set to generate a corrupted image data set comprises one or more of applying Fourier truncation, applying spatial warping, adding white noise, adding blurring, and adding coil sensitivity to the target image data set.
8. The method of claim 1 , wherein the imaging modality is ultrasound, and corrupting each target image data set to generate a corrupted image data set comprises one or more of applying spatial shading, adding white noise, adding speckle, adding blurring, and applying spatial warping to the target image data set.
9. The method of claim 1 , wherein the imaging modality is CT, and corrupting each target image data set to generate a corrupted image data set comprises one or more of adding streak artifacts, adding white noise, adding ring artifacts, and adding blurring.
10. The method of claim 1 , further comprising testing the trained neural network with image data sets acquired by the medical imaging modality.
11. A system comprising:
a memory storing a neural network; and
a processor communicably coupled to the memory and configured to:
transform each of a plurality of initial image data sets not acquired by a medical imaging modality into a target image data set, wherein each target image data set is in a format specific to the medical imaging modality;
corrupt each target image data set to generate a corrupted image data set; and
train the neural network to map each corrupted image data set to the corresponding target image data set.
12. The system of claim 11 , wherein the medical imaging modality is magnetic resonance (MR), the initial image data sets are digital photographic data sets, and transforming each initial image data set into a target image data set comprises converting color or grayscale channels of each digital photographic data set into an MR image data set of complex values.
13. The system of claim 12 , wherein the processor is further configured to add background phase to the MR image data set of complex values.
14. The system of claim 11 , wherein the medical imaging modality is MR, the initial image data sets are digital photographic data sets, and transforming each initial image data set into a target image data set comprises converting color or grayscale channels of each digital photographic data set into an MR parametric map that models one or more of proton density, relaxation time, magnetic susceptibility, chemical shift, temperature, and diffusivity.
15. The system of claim 11 , wherein the medical imaging modality is ultrasound, the initial image data sets are digital photographic data sets, and transforming each initial image data set into a target image data set comprises converting color or grayscale channels of each digital photographic data set into an ultrasound parametric map that models one or more of acoustic impedance, velocity, and density.
16. The system of claim 11 , wherein the medical imaging modality is computed tomography (CT), the initial image data sets are digital photographic data sets, and transforming each initial image data set into a target image data set comprises converting color or grayscale channels of each digital photographic data set into a CT parametric map that models radiation absorption.
17. The system of claim 11 , wherein the imaging modality is MR, and corrupting each target image data set to generate a corrupted image data set comprises one or more of applying Fourier truncation, applying spatial warping, adding white noise, adding blurring, and adding coil sensitivity to the target image data set.
18. The system of claim 11 , wherein the imaging modality is ultrasound, and corrupting each target image data set to generate a corrupted image data set comprises one or more of applying spatial shading, adding white noise, adding speckle, adding blurring, and applying spatial warping to the target image data set.
19. The system of claim 11 , wherein the imaging modality is CT, and corrupting each target image data set to generate a corrupted image data set comprises one or more of adding streak artifacts, adding white noise, adding ring artifacts, and adding blurring.
20. The system of claim 11 , wherein the processor is further configured to test the trained neural network with image data sets acquired by the medical imaging modality.