Markerless Tumor Tracking via DRR Discriminator Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Markerless tracking in radiation treatment faces challenges in accurately detecting the position of a specific site, such as a tumor, due to shape changes caused by breathing and the need for multiple templates, which reduces throughput and increases patient discomfort, and conventional methods struggle with real-time processing and accounting for region changes during treatment.
Innovation Solution
A tracking device that uses machine learning to detect the position of a specific site by creating DRR images from CT data and learning a discriminator to recognize the region, allowing for real-time tracking without the need for multiple templates and enabling accurate detection of shape changes using X-ray fluoroscopic images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple template images are stored and used for template matching to respond to shape changes at different breathing phases, then detection accuracy is improved, but treatment time increases and throughput decreases
Solution Approach 1:
The system performs preliminary action by creating DRR images from CT data and learning the discriminator model before treatment begins. This allows the machine learning model to be trained offline on synthesized images representing different breathing phases, so that during actual treatment, real-time detection can be performed quickly without requiring multiple template images to be manually prepared and matched, thus maintaining high throughput while achieving accurate detection.
2Reliability
If template matching is performed with multiple templates to account for breathing phase variations, then detection reliability is improved, but processing time increases
Solution Approach 1:
The invention replaces the mechanical template matching process with a machine learning-based discriminator system. Instead of performing computationally intensive template matching operations with multiple templates during treatment, the system uses a pre-trained neural network discriminator that can rapidly classify images and detect the specific site, significantly reducing processing time while maintaining or improving detection reliability through the learned features from diverse breathing phase representations in the DRR training data.
3Device complexity
If conventional tracking methods calculate the position of a single point in the specific site, then computational complexity is reduced, but detection precision deteriorates because the region of the specific site is not taken into account
Solution Approach 1:
The invention transitions from tracking a single point (zero-dimensional) to tracking the entire region of the specific site (two-dimensional area). The discriminator output provides pixel-wise classification across the entire image, enabling region-based tracking that captures the spatial extent and shape of the tumor. This dimensional expansion from point to region improves detection precision by accounting for the full tumor geometry while the efficient neural network architecture maintains manageable computational complexity.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Enables real-time detection and tracking of the specific site, improving treatment throughput and accuracy by eliminating the need for multiple templates and reducing redundant calculations, while ensuring precise detection even with shape changes and varying image quality.
Implementation Method 1
a DRR image creation part that creates a DRR image including the specific site by performing virtual fluoroscopic projection simulating geometrical fluoroscopic conditions between an X-ray tube and an X-ray detector with respect to the subject on CT image data
Implementation Method 2
an X-ray fluoroscopic image including the specific site of a subject... by performing discrimination using the discriminator learned by the discriminator learning part on an X-ray fluoroscopic image including the specific site
Data Source
AI summary
A control part 30 includes: a DRR image creation part 31 that creates a DRR image including a specific site; a specific site projection part 32 that creates a projection region image representing the region of the specific site; a discriminator learning part 33 that learns a discriminator for recognizing the region of the specific site by performing machine learning with use of the DRR image and the projection region image as a training label image; a specific site region detection part 34 that detects the region of the specific site by performing discrimination using the discriminator learned by the discriminator learning part 33 on an X-ray fluoroscopic image including the specific site; and a radiation signal generation part 35 that transmits a treatment beam radiation signal to an irradiation device 90 when the region of the specific site detected by the specific site region detection part 34 is included in the radiation region of a treatment beam.