Multi-label V-Net for Cardiac CTA Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for segmenting anatomical structures in cardiac CTA are computationally expensive due to the deformable registration step in multi-atlas based methods, while convolutional neural networks, although fast in deployment, are limited to binary segmentation and require optimization for multi-label tasks.
Innovation Solution
Extending the V-Net architecture to a multi-label segmentation network optimized by the sum of Dice similarity coefficients (DSC) for all anatomical structures, followed by post-processing to refine the segmentation surface.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multi-atlas based methods with deformable registration are used for segmentation, then segmentation accuracy is improved, but processing time and computational cost increase significantly
Solution Approach 1:
The patent applies preliminary action by performing deformable registration in advance to create atlases that are then reused for segmentation. The complex registration computations are done beforehand during atlas construction, allowing rapid segmentation of new images by simply matching against the pre-registered atlases, thus reducing real-time processing time while maintaining high accuracy
Solution Approach 2:
The patent segments the segmentation task into two distinct phases: atlas construction (which includes deformable registration) and atlas-based segmentation. By separating these operations, the computationally intensive registration is performed once during atlas creation, while the actual segmentation of new images uses the pre-processed atlases for faster processing
2Productivity
If convolutional neural networks are used for segmentation, then processing speed is improved, but the network is limited to binary segmentation and requires optimization for multi-label tasks
Solution Approach 1:
The patent extends the V-Net architecture to perform multi-label segmentation by modifying the network output to produce multiple segmentation maps simultaneously. The network is trained to segment multiple anatomical structures (heart chambers, great vessels, lungs) in a single forward pass, making the system universally applicable to multi-structure segmentation tasks while maintaining fast processing speeds
Solution Approach 2:
The patent changes the network parameters and loss function to optimize for multi-label segmentation. Specifically, it uses a multi-class Dice loss function and modifies the output layer to produce probability maps for multiple anatomical structures, transforming the network from binary to multi-label segmentation capability while preserving the speed advantages of CNNs
3Adaptability or versatility
If V-Net architecture is extended to multi-label segmentation, then multi-structure segmentation capability is improved, but optimization complexity increases
Solution Approach 1:
The patent implements feedback through the Dice loss function, which computes the overlap between predicted and ground truth segmentations for each anatomical structure and uses this feedback to guide gradient-based optimization. The loss function provides continuous feedback during training to adjust network parameters, enabling effective optimization of multi-label segmentation despite the increased complexity
Data Source
AI summary
A method comprises (a) collecting (i) a set of chest computed tomography angiography (CTA) images scanned in the axial view and (ii) a manual segmentation of the images, for each one of multiple organs; (b) preprocessing the images such that they share the same field of view (FOV); (c) using both the images and their manual segmentation to train a supervised deep learning segmentation network, wherein loss is determined from a multi-dice score that is the summation of the dice scores for all the multiple organs, each dice score being computed as the similarity between the manual segmentation and the output of the network for one of the organs; (d) testing a given (input) pre-processed image on the trained network, thereby obtaining segmented output of the given image; and (e) smoothing the segmented output of the given image.