Cascaded 3D DCNN for Muscle Segmentation in MRI
Solution Overview
Problem
Current methods for automatic muscle segmentation in MRI images face challenges due to GPU memory and computation speed limitations, particularly when dealing with high-resolution 3D data sets, leading to significant loss of spatial resolution and boundary information.
Innovation Solution
A cascaded 3D deep convolutional neural network (DCNN) framework that splits high-resolution 3D images into sub-images, determines boundary boxes, and crops them to segment muscles effectively, using a two-stage process to enhance segmentation accuracy and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution 3D MRI images are directly processed using DCNN, then segmentation accuracy is improved, but GPU memory requirements increase and computation speed decreases
Solution Approach 1:
The patent divides the high-resolution 3D MRI image into multiple low-resolution 3D sub-images that can be processed independently on GPU. This segmentation of the image data allows the system to maintain high segmentation accuracy through subsequent up-sampling and cropping operations while reducing the computational burden on GPU memory and processing speed
Solution Approach 2:
The patent transforms the problem from direct 3D high-resolution processing to a multi-dimensional approach: first processing low-resolution versions, then recovering high-resolution details through up-sampling and selective cropping. This dimensional transformation allows efficient GPU processing while preserving boundary information
2Productivity
If image compression is applied to fit high-resolution data into GPU memory, then computation becomes feasible, but spatial resolution and boundary information are significantly lost
Solution Approach 1:
The patent performs preliminary processing by creating low-resolution versions of the high-resolution images for initial DCNN processing. This preliminary action allows the system to identify regions of interest and crop boxes that preserve boundary information, which are then used to extract and process only the necessary high-resolution portions, avoiding the need to compress the entire high-resolution dataset
Solution Approach 2:
The patent extracts only the necessary portions of high-resolution images (cropped regions) after identifying them through low-resolution processing. This extraction approach retrieves only the relevant high-resolution data containing muscle boundaries, avoiding the information loss that would occur with comprehensive image compression
Data Source
AI summary
A method and an apparatus for automatic muscle segmentation are provided. The method includes receiving a high-resolution three-dimensional (3D) image obtained by a magnetic resonance imaging (MM) system and splitting the high-resolution 3D image into high-resolution 3D sub-images. Furthermore, the method includes acquiring low-resolution 3D sub-images of the high-resolution 3D sub-images and determining a boundary box within each corresponding low-resolution 3D sub-image. Moreover, the method includes determining a crop box for each boundary box, cropping each high-resolution 3D sub-image based on its corresponding crop box, and segmenting at least one muscle from the cropped box high-resolution 3D sub-image.