Siamese Neural Network Optical Flow for Medical Image Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current AI-based medical image segmentation methods require a large number of labeled medical image samples for training, making the process time-consuming and inefficient, especially for organ segmentation like liver segmentation in CT images.

Innovation Solution

A medical image segmentation method that uses a siamese convolutional neural network to learn optical flow motion information from labeled to unlabeled images, allowing for segmentation of target objects like organs without the need for extensive labeled data, by associating feature information and performing upsampling to determine pixel position correspondence and reconstructing target object regions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a large number of organ-labeled medical image samples are used to train the convolutional neural network, then the segmentation accuracy is improved, but the training time and processing time are significantly increased, resulting in low efficiency

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

Solution Approach 1:

The patent performs preliminary action by pre-training the convolutional neural network using a large number of labeled medical image samples offline. This pre-training establishes the basic segmentation capabilities of the network. Then, during actual application, only fine-tuning with a small number of organ-labeled samples is required, significantly reducing the training time while maintaining high segmentation accuracy. The pre-trained model serves as a foundation that requires minimal adaptation for specific organs.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If a large number of organ-labeled medical image samples are used for training, then the segmentation accuracy is improved, but the device complexity and data processing requirements are increased

Engineering Contradiction:
Improvesegmentation accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system performs preliminary training offline to establish the neural network's segmentation capabilities before deployment. This separates the complex training process from the actual segmentation application, reducing the real-time system complexity. The pre-trained model can be stored and deployed with minimal additional infrastructure requirements.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent extracts and separates the training phase from the inference phase. The time-consuming training process with large datasets is extracted and performed separately offline, while the deployment phase only requires minimal fine-tuning and actual segmentation operations. This extraction reduces the complexity of the operational system while maintaining high accuracy through the pre-trained model.

Inventive Principle:
Principle #2Taking out (Extraction)

3Adaptability or versatility

If extensive labeled data is required for training, then the model generalization performance is improved, but the labeling time and resource requirements are significantly increased

Engineering Contradiction:
Improvemodel generalization performanceVSAvoidlabeling time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent performs preliminary action by pre-training the model on diverse labeled datasets that capture various anatomical variations and imaging conditions. This pre-training provides the model with generalization capabilities across different organs and imaging scenarios. When applying to a specific organ, only a small number of labeled samples are needed for fine-tuning, dramatically reducing the labeling time while maintaining strong generalization performance.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3923238B1Medical image segmentation method and device, computer device and readable storage medium
Publication Date: 2024.10.23 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3923238B1 patent drawingFigure 1
  • EP3923238B1 patent drawingFigure 2A~2B
  • EP3923238B1 patent drawingFigure 2C~2D

AI summary

A medical image segmentation method performed by a computer device. The method comprises: acquiring a first medical image and a second medical image with a labeled target object region; performing feature extraction on the first medical image and the second medical image respectively, to obtain first feature information of the first medical image and second feature information of the second medical image; acquiring optical flow motion information of the second medical image to the first medical image according to the first feature information and the second feature information; and segmenting a target object in the first medical image according to the optical flow motion information and the labeled target object region to obtain a segmentation result of the first medical image.