Deep Neural Network Approximation for Anatomical Object Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Deep neural networks for anatomical object detection in medical images face high computational costs, making them impractical for use on devices without powerful GPUs, and existing optimization methods can degrade classification accuracy.

Innovation Solution

The method involves approximating a trained deep neural network by sparsifying weights, reducing the number of nodes, and using Haar wavelet bases and principal component analysis to reconstruct weight matrices, thereby reducing computational complexity while maintaining accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If deep neural networks are used for anatomical object detection, then classification accuracy is improved, but computational cost increases significantly

Engineering Contradiction:
Improveclassification accuracyVSAvoidcomputational cost
Core Design Contradiction:
Measurement precisionVSPower

Solution Approach 1:

The patent segments the weight matrices into multiple blocks and applies low-rank approximation to each block separately. This divides the large computational problem into smaller, more manageable pieces that can be processed efficiently while maintaining overall accuracy

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the parameter representation of weight matrices by using low-rank factorization (W ≈ UV^T) instead of storing full matrices. This parameter transformation reduces computational complexity from O(n²) to O(nr) where r << n, while preserving the essential information needed for accurate classification

Inventive Principle:
Principle #35Parameter changes

2Device complexity

If the number of nodes in deep neural network is reduced, then computational complexity is reduced, but classification accuracy may degrade

Engineering Contradiction:
Improvecomputational complexityVSAvoidclassification accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent creates compressed representations (copies) of the original weight matrices using low-rank approximation. These compressed copies use fewer parameters but capture the essential patterns needed for accurate classification, effectively copying the functional behavior with reduced complexity

Inventive Principle:
Principle #26Copying

3Productivity

If weight sparsification is applied to reduce computational complexity, then processing speed improves, but classification accuracy degrades

Engineering Contradiction:
Improveprocessing speedVSAvoidclassification accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent applies parameter changes by transforming dense weight matrices into low-rank factorized forms. This transformation enables more efficient computation while preserving the full information content, avoiding the accuracy loss that occurs with traditional sparsification methods that discard weight values

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3091486B1Method and system for approximating deep neural networks for anatomical object detection
Publication Date: 2023.08.23 SIEMENS HEALTHINEERS AG
  • EP3091486B1 patent drawingFigure 1
  • EP3091486B1 patent drawingFigure 2
  • EP3091486B1 patent drawingFigure 3

AI summary

A method and system for approximating a deep neural network for anatomical object detection is discloses. A deep neural network is trained to detect an anatomical object in medical images. An approximation of the trained deep neural network is calculated that reduces the computational complexity of the trained deep neural network. The anatomical object is detected in an input medical image of a patient using the approximation of the trained deep neural network.