Organ Segmentation Model Training Using Radiation Dose Weighting

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning models for automatic organ segmentation in radiotherapy face challenges due to inaccurate or incomplete training data, often requiring manual curation to remove poorly contoured images, which can lead to biased or poorly trained models.

Innovation Solution

A system that weights pixels or voxels based on radiation dosage applied during radiotherapy treatment to emphasize accurate areas and de-emphasize less accurate ones, allowing for the use of partially contoured images in training without discarding them.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If radiation dose information is incorporated into training data, then segmentation accuracy is improved, but data processing complexity increases

Engineering Contradiction:
Improvesegmentation accuracyVSAvoiddata processing complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The training data is segmented into multiple components: anatomical images, radiation dose information, and ground truth segmentations. Each component is processed and integrated separately through the loss function, allowing the system to handle complex multi-modal data in a structured manner that improves accuracy while managing processing complexity.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The loss function is designed to be universal and multi-functional, simultaneously handling anatomical image analysis, radiation dose integration, and segmentation evaluation in a single unified framework. This multi-functionality resolves the contradiction by providing a comprehensive solution that addresses all data types without requiring separate processing pipelines.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Reliability

If radiation dose information is used to enhance training, then model performance is improved, but computational resources required increase

Engineering Contradiction:
Improvemodel performanceVSAvoidcomputational resources
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The system applies partial action by selectively incorporating radiation dose information where it provides the most benefit for segmentation accuracy. The loss function integrates dose information in a controlled manner, using it partially rather than fully, which improves model performance while limiting the additional computational burden to only where necessary.

Inventive Principle:
Principle #16Partial or excessive action

3Manufacturing precision

If multi-modal training data including radiation dose is implemented, then segmentation precision is improved, but implementation complexity increases

Engineering Contradiction:
Improvesegmentation precisionVSAvoidimplementation complexity
Core Design Contradiction:
Manufacturing precisionVSEase of manufacture

Solution Approach 1:

The loss function serves as an intermediary that mediates between multi-modal training data (anatomical images and radiation dose information) and the segmentation model. This intermediary component simplifies implementation by providing a standardized interface for integrating diverse data types, thereby improving segmentation precision while reducing overall implementation complexity.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4315239B1Using radiation dose information for automatic organ segmentation model training
Publication Date: 2026.04.29 SIEMENS HEALTHINEERS INTERNATIONAL AG
  • EP4315239B1 patent drawingFigure 1
  • EP4315239B1 patent drawingFigure 2
  • EP4315239B1 patent drawingFigure 3A

AI summary

Disclosed herein are systems and methods for training a machine learning model for automatic organ segmentation. A processor receives 210 an image of one or more pre-contoured organs, the image comprising a plurality of voxels. The processor executes 220 a machine learning model using the image to output predicted organ labels for the plurality of voxels of the image. The processor determines 230 differences between corresponding predicted organ labels and expected organ labels for the plurality of voxels. The processor determines 240 radiation dose levels that correspond to the plurality of voxels of the image. The processor determines 250 weights for the plurality of voxels based on the radiation dose levels of the respective voxels. The processor then trains 260 the machine learning model based on the differences and the weights for the plurality of voxels.