Modular Medical Image Labeling System with Shared Backbone

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing medical image analysis models face challenges due to the scarcity of annotated training data, which limits the effectiveness of deep learning in medical imaging, especially for rare diseases or high annotation costs.

Innovation Solution

A modular labeling system that includes a shared first module for generating representations of medical images and type-specific modules for different labeling tasks, allowing for efficient multi-task training and reducing the need for redundant computations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a single deep learning model is trained for multiple labeling tasks, then the model can handle diverse medical imaging tasks, but the model may struggle with domain-specific nuances and require extensive multi-domain training data

Engineering Contradiction:
Improvemulti-task capabilityVSAvoidtask-specific performance
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent segments the model into a shared backbone network and task-specific heads. The backbone extracts general medical image features, while separate heads process each labeling task (classification, segmentation, detection) with task-specific layers, allowing each task to be optimized independently while sharing computational resources.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The shared backbone network serves multiple functions by extracting universal medical image representations that can be applied to different labeling tasks. This multi-functional component reduces redundancy while maintaining task-specific performance through separate heads.

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

2Reliability

If separate models are trained for each labeling task, then each model can be optimized for its specific task, but computational resources are wasted on redundant feature extraction

Engineering Contradiction:
Improvetask optimizationVSAvoidcomputational resource consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent merges the feature extraction functionality into a single shared backbone network that serves all tasks. Instead of training separate models with redundant feature extractors, the system combines multiple task heads that share the same backbone, reducing computational resource consumption while maintaining task optimization through specialized head layers.

Inventive Principle:
Principle #5Merging (Combining)

3Productivity

If a modular architecture with shared backbone is used, then computational resources are conserved and training efficiency improves, but the architecture becomes more complex with multiple heads

Engineering Contradiction:
Improvetraining efficiencyVSAvoidarchitecture complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The architecture is segmented into a shared backbone and separate task-specific heads. This segmentation allows for modular training where the backbone can be pre-trained and then fine-tuned with different heads for different tasks, improving training efficiency while managing complexity through standard neural network components.

Inventive Principle:
Principle #1Segmentation

4Reliability

If domain-specific pretraining is performed, then performance on that specific domain improves, but the model cannot generalize well to other medical imaging domains

Engineering Contradiction:
Improvedomain performanceVSAvoidcross-domain generalization
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system performs preliminary action by pre-training the shared backbone network on a large, diverse medical imaging dataset to learn general medical image representations. This pretraining establishes a robust foundation that can then be adapted to specific domains through task-specific heads without requiring complete retraining, thus achieving both domain performance and cross-domain generalization.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4553845A1Labeling system for labeling a medical image
Publication Date: 2025.05.14 FRAUNHOFER GESELLSCHAFT ZUR FORDERUNG DER ANGEWANDTEN FORSCHUNG EV
  • EP4553845A1 patent drawingFigure 1
  • EP4553845A1 patent drawingFigure 2
  • EP4553845A1 patent drawingFigure 3

AI summary

The invention relates to labeling of medical images using a modular model. The modules based on which the model can be generated include a) a shared first module (10) for different types of labeling (40, 41, 42), which is configured to have as input a medical image and to have as output a medical image representation (13), b) a labeling-type-specific first module (11, 12), which is specific for a type of labeling and is configured to have as input the representation provided by the shared first module and as output a labeling-type-specific representation (14, 15), and c) second modules (20, 21, 22) of which each is specific for a respective type of labeling and is configured to have as input a medical image representation and as output a labeling of the respective type of labeling. This modular structure allows for an efficient multi-task training and hence for more accurate labelings.