Deep match: A zero-shot framework for improved fiducial-free respiratory motion tracking
Di Xu1, Martina Descovich1, Hengjie Liu2
1Radiation Oncology, University of California, San Francisco, United States.
Summary
A new deep learning algorithm, Deep Match, improves fiducial-free tracking for lung stereotactic body radiation therapy (SBRT). This advancement enables more patients with small or obscured tumors to receive precise radiation treatment.
Area of Science:
- Medical Physics
- Artificial Intelligence in Medicine
- Radiation Oncology
Background:
- Lung stereotactic body radiation therapy (SBRT) requires precise motion management for effective tumor dose delivery and reduced normal tissue exposure.
- Current dynamic tracking methods include fiducial-based (invasive) and fiducial-free (challenging for small/obscured tumors) X-ray tracking.
- Existing fiducial-free tracking algorithms struggle with small tumors (<15 mm) or those obscured by surrounding anatomy.
Purpose of the Study:
- To introduce Deep Match, a novel deep learning-based template matching algorithm designed to enhance fiducial-free tracking in lung SBRT.
- To improve the accuracy and reliability of tumor tracking for cases where conventional methods fail.
Main Methods:
- Deep Match utilizes a four-stage pipeline: training-free feature extraction, similarity-based location proposal, local refinement, and uncertainty prediction.
- The algorithm was validated on a cohort of 10 patients with lung tumors that were difficult to track using conventional methods.
- Patient data was stratified by tumor size (<10 mm, 10-15 mm, >15 mm) and location (with/without thoracic overlap).
Main Results:
- Deep Match achieved >80% 3 mm-Hit accuracy for 70% of patients on X-ray views where conventional tracking failed.
- The algorithm demonstrated a superior/inferior distance (SID) with a standard deviation of approximately 1 mm across all patients.
- Robust performance was observed even for small tumors and those obscured by overlapping anatomy.
Conclusions:
- Deep Match is a zero-shot learning network that leverages deep learning without requiring patient-specific training data.
- This algorithm significantly expands the applicability of fiducial-free tracking to a broader patient population, including those with challenging tumor characteristics.
- Deep Match offers a more trustworthy and versatile solution for lung SBRT motion management.
