# US 2022/0036136 A1 · Computer-implemented method for parametrizing a function for evaluating a medical image dataset

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2022/0036136 A1
- Document ID: US20220036136A1
- Status: Published application
- Kind: A1
- Application number: 17/382,588
- Filed: 2021-07-22
- Published: 2022-02-03
- Earliest claimed: 2020-07-28
- Assignees: Siemens Healthcare GmbH
- Applicants: Siemens Healthcare GmbH
- Inventors: Alexander MUEHLBERG; Oliver TAUBMANN; Alexander KATZMANN; Felix DENZINGER; Felix LADES; Rainer KAERGEL; Felix DURLAK; Michael SUEHLING
- Web page: https://patentgrep.com/patent/US20220036136A1

## CPC

- G06F 18/217: {Validation; Performance evaluation; Active pattern learning techniques}
- G06N 3/08: Learning methods
- G06F 18/214: Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G06F 18/24765: {Rule-based classification}
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections
- G06N 3/09: Supervised learning
- G06N 3/0985: Hyperparameter optimisation; Meta-learning; Learning-to-learn
- G06V 10/25: Determination of region of interest [ROI] or a volume of interest [VOI]
- G16H 30/40: for processing medical images, e.g. editing
- G06V 2201/03: Recognition of patterns in medical or anatomical images
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G16H: HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA

## Abstract

A computer-implemented method and system are for parametrizing a function including a processing algorithm and a representation generator, the representation generator being designed to generate at least one representation. In an embodiment, the method includes using an optimization algorithm to determine the processing algorithm and the at least one representation parametrization. The optimization algorithm optimizes a measure for the performance of the processing algorithm when operating on a set of training representations generated by applying the representation generator to training medical image datasets, by varying on the one hand the content of the at least one representation parametrization and/or the number of used representation parametrizations and on the other hand the processing algorithm and the algorithm parameters.

## Claims

### Claim 1

A computer-implemented method for parametrizing a function for evaluating a medical image dataset concerning a region of interest, the function including a processing algorithm determined by the parametrization of the function and a representation generator, the representation generator being designed to generate at least one representation of the region of interest based on the medical image dataset and a respective representation parametrization determined by the parametrization of the function, and the processing algorithm being designed to determine at least one output parameter based on the at least one representation and multiple algorithm parameters, the computer-implemented method comprising:
receiving a training dataset including multiple training medical image datasets;
receiving representation information;
determining, using an optimization algorithm, the processing algorithm and the at least one representation parametrization, the optimization algorithm optimizing a measure for performance of the processing algorithm when the processing algorithm is operating on a set of training representations generated by applying the representation generator to at least a subset of the training medical image datasets using the at least one representation parametrization by varying on the one hand the content of the at least one representation parametrization based on at least one of the representation information and a number of used representation parametrizations, the processing algorithm being within a search space and the algorithm parameters; and
providing the at least one representation parametrization determined and providing the processing algorithm determined or processing parameters describing the processing algorithm determined, to parametrize the function.

### Claim 2 (depends on 1)

The computer-implemented method of claim 1, wherein the algorithm parameters are determined by the optimization algorithm or by a further optimization algorithm, wherein the further optimization algorithm optimizes the measure or a further measure for performance of the processing algorithm determined in the optimization algorithm when the processing algorithm is operating on the set or a further set of training representations generated by applying the representation generator to the training medical image datasets or to the subset or a further subset of the training medical image datasets using the at least one representation parametrization determined in the optimization algorithm by varying the algorithm parameters.

### Claim 3 (depends on 1)

The computer-implemented method of claim 1, wherein the search space defines multiple possible architectures for neural networks, wherein selection of the processing algorithm in the optimization algorithm is a result of a neural architecture search, and wherein free parameters of the neural network defined by the respective architecture, are defined by the algorithm parameters.

### Claim 4 (depends on 1)

The computer-implemented method of claim 1,
wherein the search space defines multiple convolutional neural networks that differ with respect to at least one of connections between layers and a type of operation performed by the respective layer, or
wherein the search space defines multiple neural networks that differ with respect to at least one of the connections between nodes and the activation function used in at least one of the nodes.

### Claim 5 (depends on 1)

The computer-implemented method of claim 1, wherein the representation parametrization selects between different types of representations including at least two of:
a projection image,
a multiplanar reformatted representation,
a curved planar reformation,
an unfolded representations of the region of interest,
a masked image,
a representation generated using histogram equalization, and
a normalization.

### Claim 6 (depends on 1)

The computer-implemented method of claim 1, wherein at least one of:
a possible content of the representation parametrization selects a slice of an image volume defined by the medical image dataset as the representation, the representation parametrization determining at least one of which slice of the image volume is selected and an orientation of the slice, and
a possible content of the representation parametrization selects a projection image of the image volume as the representation, the representation parametrization determining a direction of the projection.

### Claim 7 (depends on 1)

The computer-implemented method of claim 1, wherein the representation information includes or defines a set of representation parametrizations, wherein the optimization algorithm selects one of or a subgroup of the representation parametrizations in a first alternative as the at least one representation parametrization determined or in a second alternative as at least one approximate representation parametrization, and wherein the determination of the at least one representation parametrization in the second alternative is based on the at least one approximate representation parametrization.

### Claim 8 (depends on 1)

The computer-implemented method of claim 1, further comprising:
receiving user input by at least one user, the representation parametrizations defined being determined from that user input.

### Claim 9 (depends on 1)

The computer-implemented method of claim 1, wherein the optimization algorithm uses a pool of potential representation parametrizations defined by the representation information, wherein for at least a selected one of the representation parametrizations in the pool a measure of the relevance of the selected representation parametrization is determined by comparing the optimum measure for the performance achieved in the optimization algorithm when using the selected representation parametrization to generate training representations with the optimum measure for performance achieved when not using the selected representation parametrization to generate training representations.

### Claim 10 (depends on 1)

The computer-implemented method of claim 1, wherein for pixels or voxels or groups of neighboring pixels or voxels of a respective representation of the medical image dataset defined by the respective representation parametrization determined, a measure of the influence of the respective pixel or voxel or group of neighboring pixels or voxels on at least one of the output parameters, is determined based on the determined processing algorithm and algorithm parameters.

### Claim 11

A computer-implemented method for evaluating medical image data, comprising:
receiving a medical image dataset concerning a region of interest as input data;
applying a function trained by a machine learning algorithm to the medical image dataset to determine at least one output parameter, the function including a processing algorithm at least one of determined and parametrized by machine learning and a representation generator, the representation generator generating at least one representation of a region of interest based on the medical image dataset and a respective representation parametrization determined by machine learning, and the processing algorithm determining the at least one output parameter based on the at least one representation; and
providing the at least one output parameter as output data.

### Claim 12

A providing system, comprising:
a first interface, configured to receive input data;
a second interface, configured to provide output data; and
at least one processor, configured to at least one of
receive a training dataset including multiple training medical image datasets,
receive representation information,
determine, using an optimization algorithm, the processing algorithm and the at least one representation parametrization, the optimization algorithm optimizing a measure for performance of the processing algorithm when the processing algorithm is operating on a set of training representations generated by applying the representation generator to at least a subset of the training medical image datasets using the at least one representation parametrization by varying on the one hand the content of the at least one representation parametrization based on at least one of the representation information and a number of used representation parametrizations, the processing algorithm being within a search space and the algorithm parameters, and
provide the at least one representation parametrization determined and providing the processing algorithm determined or processing parameters describing the processing algorithm determined, to parametrize the function; and
receive a medical image dataset concerning a region of interest as input data,
apply a function trained by a machine learning algorithm to the medical image dataset to determine at least one output parameter, the function including a processing algorithm at least one of determined and parametrized by machine learning and a representation generator, the representation generator generating at least one representation of a region of interest based on the medical image dataset and a respective representation parametrization determined by machine learning, and the processing algorithm determining the at least one output parameter based on the at least one representation, and
providing the at least one output parameter as output data.

### Claim 13 (depends on 1)

A non-transitory computer program storing instructions which, when the computer program is executed by a computer, cause the computer to carry out the computer-implemented method of claim 1.

### Claim 14 (depends on 1)

A non-transitory computer-readable medium storing a computer program including instructions which, when the computer program is executed by a computer, cause the computer to carry out the computer-implemented method of claim 1.

### Claim 15 (depends on 3)

The computer-implemented method of claim 3, wherein the respective architecture includes input weights of nodes of the neural network.

### Claim 16 (depends on 1)

The computer-implemented method of claim 1, wherein the search space defines multiple possible architectures for neural networks, wherein selection of the processing algorithm in the optimization algorithm is a result of a neural architecture search, and wherein free parameters of the neural network defined by the respective architecture, are defined by the algorithm parameters.

### Claim 17 (depends on 2)

The computer-implemented method of claim 2,
wherein the search space defines multiple convolutional neural networks that differ with respect to at least one of connections between layers and a type of operation performed by the respective layer, or
wherein the search space defines multiple neural networks that differ with respect to at least one of the connections between nodes and the activation function used in at least one of the nodes.

### Claim 18 (depends on 2)

The computer-implemented method of claim 2, wherein at least one of
a possible content of the representation parametrization selects a slice of an image volume defined by the medical image dataset as the representation, the representation parametrization determining at least one of which slice of the image volume is selected and an orientation of the slice, and
a possible content of the representation parametrization selects a projection image of the image volume as the representation, the representation parametrization determining a direction of the projection.

### Claim 19 (depends on 11)

A non-transitory computer program storing instructions which, when the computer program is executed by a computer, cause the computer to carry out the computer-implemented method of claim 11.

### Claim 20 (depends on 11)

A non-transitory computer-readable medium storing a computer program including instructions which, when the computer program is executed by a computer, cause the computer to carry out the computer-implemented method of claim 11.
