Medical Image Segmentation via Slice Pair Feature Fusion

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current 2D convolution neural networks for segmenting medical images, such as liver images, face challenges in accurately capturing shape information due to the need to slice 3D images, leading to low segmentation accuracy as they ignore correlations between image segments during scanning.

Innovation Solution

The method involves obtaining a slice pair from a 3D medical image, performing feature extraction on each slice using different receptive fields to obtain high-level and low-level feature information, fusing this information to determine correlation between slices, and adjusting segmentation results based on the correlation relationship, thereby capturing shape information more accurately.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a 2D convolution neural network processes 3D medical images by slicing them slice by slice, then the processing complexity is reduced and the model can be trained efficiently, but the correlation between image segments is ignored and shape information of the target object cannot be captured accurately

Engineering Contradiction:
Improvemodel complexityVSAvoidsegmentation accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent transitions from processing single 2D slices to processing pairs of adjacent slices (2x2x2 volume elements) simultaneously. This dimensional extension allows the model to capture spatial correlations in the depth direction while maintaining computational feasibility, resolving the contradiction between model complexity and segmentation accuracy.

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

Solution Approach 2:

The patent merges information from multiple slices by processing slice pairs together and fusing feature maps from different slices. This combining approach preserves correlation information between adjacent slices while keeping the computational complexity manageable through shared convolutional kernels and parameter sharing mechanisms.

Inventive Principle:
Principle #5Merging (Combining)

2Productivity

If a 2D convolution neural network is used for slice-by-slice segmentation, then the training process is simplified and faster, but the segmentation accuracy is low due to ignoring correlations between slices

Engineering Contradiction:
Improvetraining speedVSAvoidsegmentation accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent segments the 3D volume into overlapping 2x2x2 sub-volumes centered at each voxel, processing them in a sliding window manner. This segmentation approach enables efficient training by breaking down the large 3D problem into smaller manageable units while ensuring complete coverage through overlaps, thus maintaining both training speed and accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

By extending the processing from single slices to slice pairs with depth information, the model captures 3D spatial correlations without significantly increasing training complexity. The use of shared convolutional kernels across the depth dimension allows efficient parameter utilization and faster convergence.

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

Data Source

PatentEP3961484B1Medical image segmentation method and device, electronic device and storage medium
Publication Date: 2024.07.17 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3961484B1 patent drawingFigure 1
  • EP3961484B1 patent drawingFigure 2~3
  • EP3961484B1 patent drawingFigure 4

AI summary

An embodiment of the present application discloses a medical image segmentation method and device, an electronic device and a storage medium; in the embodiment of the present application, after a slice pair is obtained, different receptive fields can be used to respectively perform feature extraction on each slice in the slice pair to obtain high-level feature information and low-level feature information of each slice, then, on the one hand, for each slice in the slice pair, the target object in the slice is segmented according to the low-level feature information and high-level feature information of the slice, to obtain an initial segmentation result of the slice, and on the other hand, the low-level feature information and high-level feature information of each slice in the slice pair are fused, and the association information between the slices is determined according to the fused feature information, and then, a segmentation result of the slice pair is generated according to the association information between the slices and the initial segmentation result of each slice in the slice pair; and the solution can improve the accuracy of segmentation.