Generating neural networks tailored to optimize specific medical image properties using novel loss functions
Techniques are described that facilitate generating neural network (NNs) tailored to optimize specific properties of medical images using novel loss functions. According to an embodiment, a system is provided that comprises a memory that stores computer executable components, and a processor that executes the computer executable components stored in the memory. The computer executable components comprise a training component that trains a NN to generate a modified version of computed tomography (CT) data comprising one or more optimized properties relative to the CT data using a loss function tailored to control learning adaptation of the NN based on error attributed to one or more defined components associated with the CT data, resulting in a trained NN, wherein the one or more defined components comprise at least one of a frequency component or a spatial feature component.
1 . A system, comprising:
a memory that stores computer executable components; and
a processor that executes the computer executable components stored in the memory, wherein the computer executable components comprise:
a training component that trains a neural network to generate a modified version of computed tomography data comprising one or more optimized properties relative to the computed tomography data using a loss function tailored to control learning adaptation of the neural network based on error attributed to one or more defined components associated with the computed tomography data, resulting in a trained neural network.
2 . The system of claim 1 , wherein the one or more defined components comprise at least one of a frequency component, a visual feature component, or a spatial feature component.
3 . The system of claim 1 , wherein the computed tomography data comprises sinogram data for a computed tomography image and the modified version comprises modified sinogram data for the computed tomography image.
4 . The system of claim 1 , wherein the computed tomography data comprise a computed tomography image and the modified version comprises a reconstructed computed tomography image.
5 . The system of claim 1 , wherein the computer executable components further comprise:
an inferencing component that applies the trained neural network to new computed tomography data to generate a new modified version of the new computed tomography data exhibiting the one or more optimized properties.
6 . The system of claim 5 , wherein the computed tomography data comprises sinogram data for a computed tomography image, the modified version comprises modified sinogram data for the computed tomography image, the new computed tomography data comprises new sinogram data for a new computed tomography image, the new modified version comprises new modified sinogram data for the new computed tomography image, and wherein the computer executable components further comprise:
a reconstruction component that generates a reconstructed version of the new computed tomography image using the new modified sinogram data.
7 . The system of claim 1 , wherein the one or more defined components comprise a frequency component and wherein the loss function weights the error of as a function of the frequency component.
8 . The system of claim 7 , wherein the loss function penalizes removal of defined frequencies or frequency ranges, and wherein the defined frequencies or frequency ranges are adjustable.
9 . The system of claim 1 , wherein the loss function utilizes spatial filtering of a difference between output data of the neural network and a target data set as input to the loss function.
10 . The system of claim 1 , wherein the loss function applies a penalty to a difference between input data and output data of the neural network.
11 . The system of claim 10 , wherein the loss function assesses the penalty after spatial filtering of the difference.
12 . The system of claim 9 , wherein the one or more defined components comprise a spatial feature component and wherein the training component uses a spatial feature extraction step to furnish a mask to assess the penalty on the difference.
13 . A system of claim 1 , wherein the loss function uses a composite of differences among three or more signals as input to generate a training penalty.
14 . A method, comprising:
training, by a system comprising a processor, a neural network to generate a reconstructed version of a computed tomography image comprising one or more optimized properties relative to the computed tomography image using a loss function tailored to control error in output data of the neural network based on a defined subset of variables associated with the computed tomography data; and
generating, by the system, a trained neural network model as a result of the training.
15 . The method of claim 14 , wherein the defined subset comprises at least one of, a frequency component, a visual feature component, or a spatial feature component.
16 . The method of claim 14 , wherein the computed tomography data comprises sinogram data for a computed tomography image and the modified version comprises modified sinogram data for the computed tomography image.
17 . The method of claim 14 , wherein the computed tomography data comprises a computed tomography image and the modified version comprises a reconstructed computed tomography image.
18 . The method of claim 14 , further comprising:
applying, by the system, the trained neural network to new computed tomography data to generate a new modified version of the new computed tomography data exhibiting the one or more optimized properties.
19 . The method of claim 17 , wherein the computed tomography data comprises sinogram data for a computed tomography image, the modified version comprises modified sinogram data for the computed tomography image, the new computed tomography data comprises new sinogram data for a new computed tomography image, the new modified version comprises new modified sinogram data for the new computed tomography image, and wherein the method further comprises:
generating, by the system, generates a reconstructed version of the new computed tomography image using the new modified sinogram data.
20 . The method of claim 14 , wherein the defined subset comprises a frequency variable and wherein the loss function weights the error of as a function of the frequency variable.
21 . The method of claim 14 , wherein the loss function utilizes spatial filtering of a difference between output data of the neural network and a target data set as input to the loss function.
22 . The method of claim 14 , wherein the loss function applies a penalty to a difference between input data and output data of the neural network.
23 . The method of claim 22 , wherein the loss function assesses the penalty after spatial filtering of the difference.
24 . The method of claim 22 , wherein the defined subset comprises a spatial feature component, and wherein the training comprises, performing, by the system, a spatial feature extraction step to furnish a mask to assess the penalty on the difference.
25 . A non-transitory machine-readable storage medium, comprising executable instructions that, when executed by a processor, facilitate performance of operations, comprising:
training a neural network to generate a reconstructed version of a computed tomography image comprising one or more optimized properties relative to the computed tomography image using a loss function tailored to control error in output data of the neural network based on one or more defined components, the one or more defined components comprising at least one of a frequency component or a spatial feature component; and
generating a trained neural network model as a result of the training.