# US 2026/0311339 A1 · Multiple loss functions for robust segmentation of multiple objects

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

- Number: US 2026/0311339 A1
- Document ID: US20260311339A1
- Status: Published application
- Kind: A1
- Application number: 19/490,296
- Filed: 2023-06-05
- Published: 2026-10-08
- Earliest claimed: 2023-06-05
- Applicants: IB Lab GmbH
- Inventors: Giacomo DEODATO
- Web page: https://patentgrep.com/patent/US20260311339A1

## CPC

- G06T 7/12: Edge-based segmentation
- G06N 20/00: Machine learning
- G16H 30/40: for processing medical images, e.g. editing
- G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- 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 for training a machine learning model (15), wherein the machine learning model (15) is configured to use a medical image (25) as an input for generating an output comprising at least two image segments (27, 28) of the medical image (25), wherein the training comprises optimizing a loss function measuring the predictive performance of the machine learning model (15) for a set of training data to obtain a set of optimal model parameters of the machine learning model (15) as result of the optimization, wherein the loss function comprises a loss contribution that depends on a geometrical relationship between at least two of the image segments (27, 28).

## Claims

### Claim 1

A computer-implemented method for training a machine learning model,
wherein the machine learning model is configured to use a medical image as an input for generating an output comprising at least two image segments of the medical image,
wherein the training comprises optimizing a loss function measuring the predictive performance of the machine learning model for a set of training data to obtain a set of optimal model parameters of the machine learning model as result of the optimization,
characterized in that the loss function comprises a loss contribution that depends on a geometrical relationship between at least two of the image segments.

### Claim 2 (depends on 1)

The method according to claim 1, characterized in that the machine learning model is configured to generate the at least two image segments as image masks, in particular binary masks, applicable to the medical image.

### Claim 3 (depends on 1)

The method according to claim 1, characterized in that the training data comprises a plurality of medical training images and a corresponding plurality of image training segments, wherein each medical training image comprises a graphical representation of parts of an anatomy, wherein the corresponding image training segments indicate the anatomic parts represented in the respective medical training image.

### Claim 4 (depends on 1)

The method according to claim 1, characterized in that the loss contribution depends on an overlap between at least two of the image segments.

### Claim 5 (depends on 3)

The method according to claim 3, characterized in that the geometrical relationship represents a mereology relationship between at least two of the image segments, wherein from the at least two of the image segments a first image segment represents an anatomic whole and at least one second image segment represents an anatomic part of the anatomic whole.

### Claim 6 (depends on 5)

The method according to claim 5, characterized in that the loss contribution depends on an exceedance of the first image segment by the at least one second image segment.

### Claim 7 (depends on 5)

The method according to claim 5, characterized in that the loss contribution depends on the difference between the first image segment and the at least one second image segment.

### Claim 8 (depends on 7)

The method according to claim 7, characterized in that the loss contribution is proportional to the size of the difference between the first image segment and the at least one second image segment.

### Claim 9 (depends on 5)

The method according to claim 5, characterized in that the method involves at least two machine learning models, wherein each machine learning model is separately trained using a separate loss function that comprises a loss contribution that depends on a geometrical relationship representing a mereology relationship between at least two of the output image segments, wherein the machine learning models together generate an output comprising at least three image segments, wherein each machine learning model is trained on a different level of a hierarchy of mereology relationships between the output image segments.

### Claim 10 (depends on 5)

The method according to claim 5, characterized in that from the image training segments corresponding to one medical training image a first image training segment represents an anatomic whole and at least one second image training segment represents an anatomic part of the anatomic whole, wherein the loss contribution further depends on the difference between a first image training segment and the at least one second image training segment.

### Claim 11 (depends on 1)

A method for segmentation of anatomical structures in a medical image comprising the steps of:
training a machine learning model to obtain a set of optimal model parameters according to claim 1;
segmenting anatomical structures in said medical image by using said machine learning model with said optimal model parameters.

### Claim 12 (depends on 11)

The method according to claim 11, wherein the medical image comprises a spine, a hand, a foot, a cell or an embryo.

### Claim 13 (depends on 1)

A set of optimal model parameters for a machine learning model, wherein the machine learning model is configured to use a medical image as an input for generating an output comprising at least two image segments of the medical image, characterized in that the set of optimal model parameters has been obtained according to claim 1.

### Claim 14 (depends on 13)

A computer program for segmentation of anatomical structures in a medical image, wherein the computer program comprises the set of optimal model parameters and the machine learning model according to claim 13 and instructions to cause a data processing apparatus to execute the step of segmenting anatomical structures in said medical image by using said machine learning model with said optimal model parameters.

### Claim 15 (depends on 14)

A computer-readable medium having stored thereon the computer program of claim 14.
