Adaptive Radiotherapy Auto-Contouring via Shape Dictionary
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current radiotherapy methods for auto-contouring in medical images, particularly for organs like the prostate, face challenges such as low contrast, image artifacts, and unpredictable organ conditions, leading to inefficiencies and inaccuracies in contouring and treatment planning.
Innovation Solution
A learning-based system that uses a population-based boundary detector and sparse shape dictionary to refine contours in medical images, allowing for adaptive and efficient auto-contouring by updating the detector and dictionary with patient-specific information, enabling accurate and consistent contouring even in low-quality images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deformable registration is employed to improve contour accuracy by calculating non-linear organ deformations, then measurement precision is improved, but computational cost increases in proportion to the number of reference images used
Solution Approach 1:
The system pre-trains a boundary detector and shape dictionary offline using a large number of reference images and patient data. This preliminary action captures organ shape variations and boundary characteristics in advance, so that during online adaptive planning, the pre-trained models can be applied directly without requiring extensive real-time computation with multiple reference images, thus reducing computational cost while maintaining high contour accuracy
Solution Approach 2:
The system creates a shape dictionary that stores representative organ shape models learned from multiple reference images. Instead of performing deformable registration with each reference image during online planning, the system copies and adapts the pre-learned shape models to the current patient anatomy, significantly reducing computational requirements while preserving the accuracy benefits of using multiple reference images
2Measurement precision
If multiple reference images are used to improve contouring accuracy, then measurement precision is improved, but productivity decreases due to increased computational cost
Solution Approach 1:
The boundary detector and shape dictionary are trained offline in advance using large datasets of reference images and manual contours. This preliminary training phase captures organ shape variations and boundary characteristics, enabling the system to perform fast online contouring without requiring extensive real-time computation, thus improving both accuracy and productivity
Solution Approach 2:
The system replaces the traditional mechanical deformable registration process with a learning-based approach using pre-trained boundary detectors and shape dictionaries. This substitution transforms the computationally intensive registration process into a faster inference process, significantly improving contouring speed while maintaining accuracy
3Device complexity
If rigid registration is used to map previous contours to current images, then device complexity is reduced, but measurement precision deteriorates due to limited degrees of freedom
Solution Approach 1:
The system uses a pre-trained boundary detector that has learned organ shape characteristics from multiple reference images. Instead of relying on simple rigid registration, the system copies and adapts these learned shape models to the current image, achieving higher accuracy without the complexity of full deformable registration
Solution Approach 2:
The system transforms the registration problem into a shape model adaptation problem. By changing from rigid transformation parameters to learned shape dictionary parameters, the system achieves better accuracy while maintaining computational efficiency, as the shape dictionary encodes complex organ variations in a compact form
Data Source
Figure 1(a)~1(e)
Figure 2
Figure 3A~3C
AI summary
A method for performing automatic contouring in a medical image. The method may include receiving an image containing a region of interest and determining a first contour of the region of interest using a boundary detector. The method may include refining the first contour based on a shape dictionary to generate a second contour of the region of interest and updating at least one of the boundary detector or the shape dictionary based on the second contour.