Medical image enhancement using an artificial intelligence model with editable output image appearance control
Medical image enhancement processing techniques are described that employ an artificial intelligence (AI) model with editable output image appearance control. An example, computer-implemented method comprises generating a transformed version of a medical image via execution of an AI transformation model on the medical image, wherein the AI transformation model comprises a neural network that predicts values of parameters of a transformation function based on processing the medical image, or a down sampled version thereof, via the neural network, and a transformation module that generates the transformation function using the values and applies the transformation function to the medical image, resulting in generation of the transformed version of the medical image. The AI transformation model comprises an editable output image appearance control functionality that enables a user to control and edit the transformation operations performed by the transformation function on the medical image in association with viewing the results in real-time.
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:
an execution component that generates a transformed version of a medical image via execution of a transformation model on the medical image, wherein the transformation model comprises:
a neural network that processes the medical image or a down sampled version of the medical image to predict values of parameters of a transformation function; and
a transformation module that configures the transformation function using the values, resulting in a first configured version of the transformation function, and applies the first configured version of the transformation function to the medical image, to generate the transformed version of the medical image; and
a rendering component that renders the transformed version of the medical image on an electronic display via a graphical user interface.
2 . The system of claim 1 , wherein the graphical user interface comprises an editing tool that facilitates receiving user input indicating an adjustment to one or more of the values that controls an appearance of the transformed version,
wherein in response to reception of the user input, the transformation module updates the transformation function in accordance with the adjustment, resulting in a second configured version of the transformation function, and applies the second configured version of the transformation function to the medical image to generate an updated transformed version of the medical image without reapplying the neural network to the medical image or the down sampled version of the medical image, and
wherein the rendering component renders the updated transformed version of the medical image on the electronic display via the graphical user interface.
3 . The system of claim 2 , wherein the editing tool comprises interactive parameter control information, and wherein the editing tool facilitates receiving the user input in association with adjusting the interactive parameter control information via the graphical user interface.
4 . The system of claim 2 , wherein the editing tool comprises an interactive graphical representation of the transformation function, and wherein the editing tool facilitates receiving the user input via the interactive graphical representation.
5 . The system of claim 1 , wherein the transformation function comprises mapping information defining a mapping between input pixel intensities of respective pixels of the medical image and output pixel intensities for corresponding pixels of the transformed version, wherein the transformation module generates the mapping information in accordance with the values and predefined relationships between the parameters, and wherein the transformation module adjusts the input pixel intensities in accordance with the mapping information, resulting in the generation of the transformed version of the medical image.
6 . The system of claim 5 , wherein the mapping information corresponds to a graphical look-up curve.
7 . The system of claim 5 , wherein the values comprise per pixel values of the parameters for each pixel of the respective pixels.
8 . The system of claim 1 , wherein the neural network comprises a combination of a convolutional neural network encoder and regression layers and excludes a decoder neural network, wherein the transformation function comprises a pixel intensity transformation function, and wherein the transformed version of the medical image comprises a pixel intensity transformed version of the medical image without artifacts as a result of the neural network excluding the decoder neural network.
9 . The system of claim 1 , wherein the computer-executable components further comprise:
a training component that trains the neural network based on a training dataset, wherein the training dataset includes training medical images and ground-truth transformed versions of the training medical images.
10 . The system of claim 9 , wherein the training component trains the transformation model using a training process that comprises, for each training medical image:
predicting, via the neural network, training values of the parameters of the transformation function based on processing the training medical image or a down sampled version of the training medical image via the neural network;
generating, via the transformation module, a tailored version of the transformation function for the training medical image using the training values;
applying, via the transformation module, the tailored version of the transformation function to the training medical image, resulting in generation of a training transformed version of the training medical image; and
tuning, by the system, network parameters of the neural network based on a measure of loss between the training transformed version and a corresponding ground-truth transformed version of the training medical image.
11 . The system of claim 1 , wherein the transformation function comprises a combination of two or more different transformation functions.
12 . A method, comprising:
applying, by a system operatively coupled to a processor, a neural network model to a medical image or a down sampled version of the medical image to predict values of parameters of a transformation function;
configuring, by the system, the transformation function using the values, resulting in a first configured version of the transformation function;
applying, by the system, the first configured version of the transformation function to the medical image to generate the transformed version of the medical image; and
rendering, by the system, the transformed version of the medical image on an electronic display via a graphical user interface.
13 . The method of claim 12 , further comprising:
providing, by the system via the graphical user interface, an editing tool that facilitates receiving user input indicating an adjustment to one or more of the values that controls an appearance of the transformed version;
updating, by the system in response to reception of the user input, the first configured version of the transformation function in accordance with the adjustment, resulting in a second configured version of the transformation function;
applying, by the system, the second configured version of the transformation function to the medical image to generate an updated transformed version of the medical image without reapplying the neural network to the medical image or the down sampled version of the medical image; and
rendering, by the system, the updated transformed version of the medical image on the electronic display via the graphical user interface.
14 . The method of claim 13 , wherein the editing tool comprises interactive parameter control information, and wherein the editing tool facilitates receiving the user input in association with adjusting the interactive parameter control information via the graphical user interface.
15 . The method of claim 12 , wherein the first configured version of the transformation function comprises mapping information defining a mapping between input pixel intensities of respective pixels of the medical image and output pixel intensities for corresponding pixels of the transformed version, and wherein configuring the transformation function comprises generating the mapping information in accordance with the values and predefined relationships between the parameters.
16 . The method of claim 15 , wherein the applying the first configured version of the transformation function comprises adjusting the input pixel intensities in accordance with the mapping information, and wherein the mapping information corresponds to a graphical look-up curve.
17 . The method of claim 12 , further comprising:
training, by the system, the neural network model based on a training dataset, wherein the training dataset includes training medical images and ground-truth transformed versions of the training medical images.
18 . The method of claim 17 , wherein the training comprises, for each training medical image:
predicting, via the neural network model, training values of the parameters of the transformation function based on processing the training medical image or a down sampled version of the training medical image via the neural network;
generating, via the transformation module, a tailored version of the transformation function for the training medical image using the training values;
applying, via the transformation module, the tailored version of the transformation function to the training medical image, resulting in generation of a training transformed version of the training medical image; and
tuning, by the system, network parameters of the neural network based on a measure of loss between the training transformed version and a corresponding ground-truth transformed version of the training medical image.
19 . A non-transitory machine-readable storage medium, comprising executable instructions that, when executed by a processor, facilitate performance of operations, comprising:
generating a transformed version of a medical image via execution of a transformation model on the medical image, wherein the transformation model comprises a neural network model and a transformation function;
rendering the transformed version of the medical image on an electronic display;
providing, via a graphical user interface rendered on the electronic display, an editing tool that facilitates receiving user input indicating an adjustment to one or more values of one or more parameters of the transformation function that controls an appearance of the transformed version;
in response to reception of the user input, updating the transformation function in accordance with the adjustment, resulting in an updated version of the transformation function;
applying the updated version of the transformation function to the medical image to generate an updated transformed version of the medical image without applying the neural network model to the medical image; and
rendering the updated transformed version of the medical image on the electronic display.
20 . The non-transitory machine-readable storage medium of claim 19 , wherein the neural network comprises a combination of a convolutional neural network encoder and regression layers and excludes a decoder neural network, wherein the transformation function comprises a pixel intensity transformation function, and wherein the transformed version of the medical image comprises a pixel intensity transformed version of the medical image without artifacts as a result of the neural network excluding the decoder neural network.