Brain Image Segmentation via Multi-Modal Feature Fusion

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional brain image segmentation methods using deep learning techniques often discard information from multiple modal images during fusion, leading to limited feature expression and reduced segmentation accuracy.

Innovation Solution

A method that involves obtaining multiple modal images of the brain, performing skull stripping, extracting features from each image separately, and fusing them to obtain a fusion feature for initial segmentation, followed by masking to remove false positives, thereby improving feature expression and segmentation accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If multiple modal images are fused during skull stripping, then skull stripping efficiency is improved, but information from some modes is discarded leading to limited feature expression

Engineering Contradiction:
Improveskull stripping efficiencyVSAvoidinformation from modal images
Core Design Contradiction:
ProductivityVSLoss of information

Solution Approach 1:

The patent segments the processing into two distinct phases: skull stripping phase (using only T1 and T1_IR images) and segmentation phase (using all three modal images including T2). This segmentation allows efficient skull stripping without information loss, while preserving complete information for the subsequent segmentation task.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs skull stripping as a preliminary action before full segmentation. By completing skull stripping using a subset of images first, the method prepares the data structure for subsequent comprehensive feature extraction and segmentation using all available modal images.

Inventive Principle:
Principle #10Preliminary action

2Device complexity

If features are extracted from fused modal images, then processing is simplified, but feature expression capability is limited

Engineering Contradiction:
Improveprocessing complexityVSAvoidfeature expression capability
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The patent segments feature extraction into two stages: first extracting features from T1 and T1_IR for skull stripping, then extracting features from all three modal images (T1, T1_IR, T2) for segmentation. This segmented approach maintains processing simplicity while maximizing feature expression capability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies different processing quality to different processing stages: simplified feature extraction for skull stripping, and comprehensive multi-modal feature extraction for segmentation. This local quality differentiation optimizes both processing efficiency and feature expression capability at appropriate stages.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP3916674B1Brain image segmentation method, apparatus, network device and storage medium
Publication Date: 2024.04.10 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3916674B1 patent drawingFigure 1~2
  • EP3916674B1 patent drawingFigure 3
  • EP3916674B1 patent drawingFigure 4

AI summary

Disclosed by embodiments of the present application are a brain image segmentation method and apparatus, a network device, and a storage medium; specifically, after acquiring an image set to be segmented, on one hand, skull peeling may be performed according to multiple modal images among the image set to be segmented so as to obtain a skull peeling mask; on the other hand, feature extraction and fusion may be performed on the multiple modal images respectively, and then intracranial tissue is segmented according to the fused features; then, an initial segmentation result obtained by means of segmentation and a previously obtained mask are fused to obtain a final segmentation result.