DCNN Image Segmentation for Radiotherapy Planning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Manual segmentation of medical images for radiotherapy planning is time-consuming and prone to variations, while fully automated techniques face challenges due to image noise and limited contrast, especially for soft-tissue structures.
Innovation Solution
A multi-model deep convolutional neural network (DCNN) scheme is employed, where a 2D or 2.5D model is used to define a region-of-interest within the 3D image, allowing a separate 3D DCNN to provide enhanced segmentation detail without processing the entire 3D volume, leveraging the outputs from both models for improved accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a 3D DCNN model is applied to process the full 3D image data set, then segmentation accuracy is improved, but segmentation duration becomes unacceptably long
Solution Approach 1:
The patent divides the full 3D image dataset processing into two stages: first, a 2D/2.5D model processes the entire volume to generate an initial segmentation and identify a region-of-interest; second, a 3D DCNN model processes only the extracted ROI to refine the segmentation. This segmentation of the processing task itself resolves the contradiction by applying the computationally intensive 3D model only where needed rather than to the entire dataset.
Solution Approach 2:
The patent performs preliminary processing using a 2D/2.5D DCNN model to generate an initial segmentation and automatically define the region-of-interest before applying the 3D DCNN model. This preliminary action identifies and isolates the relevant area, allowing the subsequent 3D processing to focus computational resources on a smaller, more manageable subset of the data, thereby reducing overall processing time while maintaining accuracy.
2Measurement precision
If manual segmentation is performed, then segmentation accuracy can be maintained, but time consumption increases significantly
Solution Approach 1:
The patent implements an automated system where the 2D/2.5D model performs initial segmentation and automatically defines the region-of-interest for the 3D model, eliminating the need for manual intervention in ROI selection. The system serves itself by using its own output (initial segmentation) to guide subsequent processing, thereby maintaining high accuracy while dramatically improving productivity compared to manual segmentation.
3Adaptability or versatility
If atlas-based auto-segmentation is used, then prior anatomical information is incorporated, but segmentation accuracy is unsatisfactory for certain applications
Solution Approach 1:
The patent replaces the mechanical atlas-based matching approach with a deep learning-based DCNN system. Instead of relying on pre-defined anatomical atlases and manual or automated registration processes, the DCNN models learn anatomical patterns directly from training data and automatically perform segmentation, thereby improving accuracy while maintaining the ability to incorporate anatomical information.
Data Source
AI summary
Features, such as anatomical features, may be automatically segmented from medical imaging information, using a computer-implemented method. In an example, three-dimensional (3D) medical imaging information may be received, such as defining a first volume. A first trained convolutional neural network (CNN) may be applied to the three-dimensional medical imaging information. An output from the first trained CNN may be used to determine a region-of-interest within the first volume, the region-of-interest defining a lesser, second volume. A different, second trained CNN may be applied to the region-of-interest, a segmented representation of the 3D medical imaging information may be provided using the outputs from the first and second CNNs, where the second CNN provides enhanced segmentation detail in the region-of-interest without requiring application of the second CNN to an entirety of the first volume. Techniques are also described from training one or more of the first and second CNNs.