Medical Image Landmark Detection with Consultant-Guided Agents

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for automated landmark detection in medical images face challenges related to performance, robustness, reliability, and efficiency, particularly in regions with little or no image data, leading to agents becoming 'stuck' and unable to detect landmarks reliably.

Innovation Solution

A multi-agent deep reinforcement learning approach combined with a consultant module neural network is used, where each agent calculates initial action values and a consultant module determines additional action values based on all agents' data, optimizing movement towards landmarks, and the network is trained separately to reduce memory burden.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If multiple agents are trained to detect landmarks independently using reinforcement learning, then each agent can specialize in specific landmark types, but the system becomes less reliable in regions with little or no image data where agents may become stuck

Engineering Contradiction:
Improvelandmark detection capabilityVSAvoiddetection reliability
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

A consultant module is introduced as an intermediary component that receives image data and provides guidance to multiple specialized agents. The consultant module processes the image data and generates guidance information that helps agents navigate and detect landmarks, particularly in data-sparse regions where independent agents would otherwise become stuck. This mediator coordinates the specialized agents' efforts while providing the reliability needed in challenging imaging conditions.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If a centralized multi-agent system is used to coordinate landmark detection, then reliability improves through collaboration, but computational complexity and memory requirements increase

Engineering Contradiction:
Improvedetection reliabilityVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system is segmented into distinct functional components: a consultant module that processes image data and generates guidance, and multiple specialized agents that execute detection tasks. Each component has a specific role and processes data independently to the extent possible, reducing the complexity of coordination while maintaining collaborative benefits. The consultant module handles the complex image analysis, while agents focus on action selection based on provided guidance.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The complex task of image analysis and landmark identification is extracted from the individual agents and placed in the consultant module. This extraction allows agents to remain relatively simple decision-making units that specialize in specific landmark types, while the consultant module handles the computationally intensive image processing and coordination functions.

Inventive Principle:
Principle #2Taking out (Extraction)

3Productivity

If agents are trained separately to reduce memory burden, then training efficiency improves, but coordination between agents becomes more difficult

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

Solution Approach 1:

The consultant module serves as a mediator that all agents interact with during both training and operation. This standardized interface simplifies coordination by providing a common communication channel and consistent guidance format, even though agents are trained separately. The consultant module's guidance output acts as a universal language that all specialized agents can understand and act upon.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3885990B1Landmark detection
Publication Date: 2025.07.02 SIEMENS HEALTHINEERS AG
  • EP3885990B1 patent drawingFigure 1
  • EP3885990B1 patent drawingFigure 2~3
  • EP3885990B1 patent drawingFigure 4~5

AI summary

A computer implemented method for detecting one or more landmarks of a medical image, and apparatus for performing the method, is disclosed. A trained multi-agent module applies each of a plurality of agents to a respective one of a plurality of sub-regions of image data to calculate a plurality of sets of first action values. A trained consultant module neural network is applied to the plurality of sets of first action values to determine one or more second action values. Each of the plurality of agents receive the one or more second action values determined for the agent by the trained consultant module neural network, and determine, based on the obtained one or more second action values and the set of first action values calculated by the agent, an action to be performed by the agent. Each agent performs the determined action thereby to move the associated sub-region.