Medical Image Model Training Using Segmentation Guidance

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Trained machine learning models in medical image processing lack robustness and explainability, often prioritizing irrelevant features due to insufficient similarity with training data, leading to incorrect decisions and hard-to-detect errors, especially in classification tasks and image-to-image processing.

Innovation Solution

Integrate segmentation information from established algorithms into the training process by using a segmentation algorithm to determine segmentation information for training datasets, incorporating segment-specific terms into the loss function to guide model training, and generate explanatory images to highlight relevant features.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a trained machine learning model is used for medical image processing, then classification accuracy and image processing capability are improved, but robustness and explainability deteriorate due to the black box nature of the model

Engineering Contradiction:
Improveclassification accuracyVSAvoidmodel robustness and explainability
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent introduces segmentation information as an intermediary element that mediates between the input medical image and the machine learning model. This segmentation information explicitly identifies and segments relevant features in the medical image, providing a structured representation that guides the model's attention. By using this intermediary segmentation layer, the model achieves improved robustness and explainability while maintaining classification accuracy, as the segmentation acts as a bridge that makes the model's decision-making process more transparent and reliable.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If the machine learning model is trained with standard supervised learning, then the model can process images, but it prioritizes irrelevant features when input data is not sufficiently similar to training data

Engineering Contradiction:
Improveimage processing capabilityVSAvoidfeature selection accuracy
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent applies preliminary action by performing segmentation of the medical image before feeding it to the machine learning model. The segmentation process pre-identifies and marks relevant features (such as lesions, tumors, or anatomical structures) before the main classification task. This preliminary segmentation step ensures that when the model processes images that differ from training data, it still prioritizes the correctly identified relevant features rather than irrelevant ones, thereby improving feature selection accuracy and reliability.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If a large number of training datasets is used to minimize errors, then model accuracy improves, but training time and computational resources increase

Engineering Contradiction:
Improvemodel accuracyVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent extracts and utilizes segmentation information from medical images to create a more efficient training approach. By extracting segmentation masks that highlight relevant features, the training process can focus on these segmented regions rather than processing entire images. This extraction of key information allows for reduced training dataset requirements and shorter training times while maintaining or improving model accuracy, as the model learns from concentrated, relevant feature representations rather than raw pixel data.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP4712018A1Computer-implemented method for providing a trained machine learning model and for evaluating an input medical image, data processing system, medical imaging device, computer program and computer-readable medium
Publication Date: 2026.03.18 SIEMENS HEALTHINEERS AG
  • EP4712018A1 patent drawingFigure 1
  • EP4712018A1 patent drawingFigure 2~3
  • EP4712018A1 patent drawingFigure 4~5

AI summary

Computer-implemented method for providing a trained machine learning model (22) that is designed to provide output data (23) based on an input medical image (24) supplied to the trained machine learning model (22), the method comprising the steps of: - receiving multiple training datasets (25), wherein the respective training dataset (25) consists of or comprises a respective training medical image (26), - applying a segmentation algorithm (27) to the respective training medical image (26) to determine a respective segmentation information (28) for the respective training dataset (25), wherein the respective segmentation information (28) concerns at least one segment (29-31) of the respective training medical image (26), - training a machine learning model (22) based on the training datasets (25) and the respective segmentation information (28) for the respective training dataset (25), and - providing the trained machine learning model (22).