Marginal Space Deep Neural Networks for Anatomical Pose Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing anatomical object detection methods in medical images, such as those using marginal space learning (MSL), are not robust when anatomical objects exhibit large variations in anatomy, shape, or appearance, particularly in complex medical imaging modalities like CT, MRI, and ultrasound, due to the reliance on handcrafted image features which are inefficient in capturing complex distributions.

Innovation Solution

The use of deep neural networks that directly input image patches from training data to learn high-level domain-specific features, dividing the parameter space into marginal search spaces of increasing dimensionality for more accurate anatomical object detection, employing discriminative and regression functions to calculate probabilities and displacement vectors for pose estimation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If handcrafted image features are used for anatomical object detection, then the detection process is simpler, but the robustness and accuracy deteriorate when anatomical objects exhibit large variations in anatomy, shape, or appearance

Engineering Contradiction:
Improvedetection process complexityVSAvoiddetection robustness
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The patent replaces handcrafted feature extraction methods with deep neural networks that automatically learn features from raw image data. This substitution enables the system to capture complex anatomical variations and patterns that handcrafted features cannot represent, significantly improving detection robustness while maintaining computational feasibility through efficient network architectures.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent transforms the feature representation parameters from fixed handcrafted descriptors to adaptive learned features through deep neural networks. By changing the parameter space from predefined features to data-driven feature representations, the system can dynamically adapt to various anatomical variations, improving reliability without excessive complexity.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If deep neural networks are used to learn high-level features directly from image patches, then detection accuracy improves, but computational complexity increases

Engineering Contradiction:
Improvedetection accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the parameter space into multiple marginal search spaces with increasing dimensionality. Instead of searching the entire high-dimensional space at once, the system divides it into manageable subsets (e.g., position, orientation, scale separated), training separate deep neural networks for each marginal space. This segmentation reduces computational complexity while maintaining high detection accuracy through focused learning in each subspace.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transforms the high-dimensional parameter space search problem into a sequence of lower-dimensional marginal space problems. By changing from a single high-dimensional search to multiple low-dimensional searches, the system reduces computational burden while preserving detection accuracy through the hierarchical structure of marginal spaces.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Measurement precision

If the parameter space is divided into marginal search spaces with increasing dimensionality, then detection precision improves, but the number of networks to train increases

Engineering Contradiction:
Improvepose estimation precisionVSAvoidnumber of networks
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the parameter space into meaningful marginal spaces (position, orientation, scale) and trains dedicated networks for each. This segmentation allows precise estimation of each parameter by specialized networks, improving overall pose estimation precision. The modular structure also enables efficient training and deployment compared to a single monolithic network.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Each deep neural network in the marginal space framework is designed to be universal within its specific search space, learning features that are transferable across different anatomical structures and imaging modalities. This universality reduces the need for extensive customization and retraining, mitigating the complexity increase from having multiple networks.

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

Data Source

PatentEP3093821B1Method and system for anatomical object pose detection using marginal space deep neural networks
Publication Date: 2019.10.09 SIEMENS HEALTHCARE GMBH
  • EP3093821B1 patent drawingFigure 1
  • EP3093821B1 patent drawingFigure 2
  • EP3093821B1 patent drawingFigure 3

AI summary

A method and system for anatomical object detection using marginal space deep neural networks is disclosed. The pose parameter space for an anatomical object is divided into a series of marginal search spaces with increasing dimensionality. A respective deep neural network is trained for each of the marginal search spaces, resulting in a series of trained deep neural networks. Each of the trained deep neural networks can evaluate hypotheses in a current parameter space using discriminative classification or a regression function. An anatomical object is detected in a medical image by sequentially applying the series of trained deep neural networks to the medical image.