3D Medical Image Segmentation via Overlapping Sub-Image CNNs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical image segmentation methods using neural networks are often non-specific to medical imaging and do not effectively leverage the structured and recurrent nature of anatomical features in medical images, leading to inefficiencies in identifying and segmenting anatomical structures.
Innovation Solution
The method involves dividing a 3D medical image into nine partially overlapping sub-images, each analyzed by a convolutional neural network (CNN), and combining their results to achieve a single segmentation of the initial image, utilizing a knowledge database to exploit the specific localization of organs and tissues.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single neural network analyzes the entire 3D medical image, then the segmentation process is simple, but the processing time is long and efficiency is low
Solution Approach 1:
The patent divides the 3D medical image into multiple 2D slices along the Z-axis, with each slice further divided into four quadrants. This segmentation allows parallel processing of multiple regions by different neural networks, significantly reducing processing time while maintaining segmentation accuracy through the structured division of the image into manageable portions.
Solution Approach 2:
The patent transforms the 3D image analysis problem into a series of 2D slice analyses. By processing each 2D slice independently and then combining the results, the system achieves efficient parallel processing while maintaining the 3D anatomical context. This dimensional transformation enables faster processing compared to analyzing the entire 3D volume with a single network.
2Measurement precision
If standard neural network architecture is used, then the model is simple and easy to implement, but it lacks specificity to medical imaging content
Solution Approach 1:
The patent assigns different neural networks to different quadrants of each 2D slice, allowing each network to be specialized for specific anatomical regions. This local quality approach enables each network to develop region-specific features and patterns, improving segmentation accuracy for that particular anatomical area while maintaining overall system functionality.
Solution Approach 2:
The patent creates a universal framework where multiple neural networks process different parts of the same image simultaneously. Each network follows the same basic architecture and training methodology but is applied to different spatial regions, achieving both specialization for specific anatomical structures and generalization through the unified processing framework.
3Loss of time
If the image is divided into multiple sub-images for parallel processing, then processing time is reduced, but the complexity of combining results increases
Solution Approach 1:
The patent merges the segmentation results from multiple 2D slices and quadrants by combining the labeled pixel information across all sub-images. The merging process reconstructs the complete 3D segmentation by aggregating results from all processed regions, maintaining anatomical consistency through the structured combination approach that leverages the systematic division of the original image.
Data Source
Figure 1
Figure 2~3B
Figure 4~5
AI summary
This invention concerns an automatic segmentation method of a medical image making use of a knowledge database containing information about the anatomical and pathological structures or instruments, that can be seen in a 3D medical image of a x b x n dimension, i.e. composed of n different 2D images each of a x b dimension. Said method being characterised in that it mainly comprises three process steps, namely: a first step consisting in extracting from said medical image nine sub-images (1 to 9) of a/2 x b/2 x n dimensions, i.e. nine partially overlapping a/2 x b/2 sub-images from each 2D image; a second step consisting in nine convolutional neural networks (CNNs) analysing and segmenting each one of these nine sub-images (1 to 9) of each 2D image; a third step consisting in combining the results of the nine analyses and segmentations of the n different 2D images, and therefore of the nine segmented sub-images with a/2 x b/2 x n dimensions, into a single image with a x b x n dimension, corresponding to a single segmentation of the initial medical image.