Tumor Tracking via 3D Segmentation and Graph Propagation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for tracking tumors in sequences of images are complex and sensitive to the selection of sink and source pixels, making them inefficient for real-time tumor tracking in particle beam radiotherapy, especially when the tumor moves or becomes invisible.
Innovation Solution
The method formulates tumor tracking as a time-wise progressive 3D segmentation problem using a 3D model of the tumor as prior information, incorporating bi-plane images and graph propagation techniques to infer the tumor's location and volume, with particle filtering and maximum a posteriori estimation to refine segmentation hypotheses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If graph cut based segmentation techniques are used for tumor tracking, then segmentation accuracy is improved, but computational time increases making real-time tracking difficult
Solution Approach 1:
The patent applies preliminary action by pre-processing the ultrasound images to enhance tumor visibility and pre-segmenting the image sequence into frames. By preparing the data in advance and identifying candidate tumor regions before applying the full graph cut algorithm, the computational burden is reduced while maintaining segmentation accuracy.
Solution Approach 2:
The patent segments the tumor tracking problem into multiple stages: initial tumor detection, candidate region identification, and refined segmentation using graph cuts. This multi-stage segmentation approach allows the computationally intensive graph cut algorithm to be applied only to relevant regions rather than entire images, reducing overall computational time.
2Measurement precision
If level set techniques are used for tracking tumor boundaries, then boundary detection accuracy is improved, but the method becomes sensitive to initial conditions and computationally intensive
Solution Approach 1:
The patent introduces an intermediary step between level set evolution and final segmentation by using graph cut algorithms as a mediator. The level set method provides an initial boundary estimate, which then serves as input for the graph cut refinement process, combining the strengths of both methods while mitigating their individual weaknesses.
Solution Approach 2:
The patent uses multiple copies of the tumor boundary at different evolution stages of the level set method. By maintaining and comparing multiple boundary representations, the system can select the most accurate boundary while reducing sensitivity to initial conditions through ensemble evaluation.
3Reliability
If the tumor becomes invisible or moves significantly between frames, then tracking reliability deteriorates, but maintaining continuous tracking is critical for real-time radiotherapy
Solution Approach 1:
The patent applies beforehand cushioning by maintaining a historical database of tumor positions and characteristics from previous frames. When the tumor becomes invisible or moves significantly, this historical information serves as a cushion that allows the tracking system to predict and recover the tumor location, preventing complete tracking failure.
Solution Approach 2:
The patent implements feedback mechanisms where the segmentation results from each frame are fed back into the next frame's processing. This continuous feedback loop allows the system to adapt to tumor motion and disappearance by using recent historical data to guide the search and segmentation process in subsequent frames, maintaining tracking reliability despite challenging conditions.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A tumor is tracked in sequences of bi-plane images by generating a set of segmentation hypotheses using a 3D model of the tumor, a bi-plane geometry, and a previous location of the tumor as determined from the pairs of biplane images. Volume prior probabilities are constructed based on the set of hypotheses. Seed pixels are selected using the volume prior probabilities, and a bi-plane dual image graph is constructed using intensity gradients and the seed pixels to obtaining segmentation masks corresponding to tumor boundaries using the image intensities to determine a current location of the tumor.