Target Motion Tracking with LLE-Based Features Using Surface Imaging Model
Presentations
PO-GePV-M-107 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: We propose to identify the target location inside human body based on features acquired from simple surface imaging with 4DCT and LLE (Locally Linear Embedding).
Methods: Our model is based on two 10-phase lung 4DCTs taken from the same patient within half-hour interval. We labeled targets in 4DCT, and then assigned 10 target positions initially corresponding to each phase of 4DCT. The first 4DCT is for training, and second is for testing. Firstly, a bounding box containing the anterior breast region of patient is contoured. Secondly, the outer skin surface of chest region is transformed into a mask, and the mask is extended into a single column vector with a high dimension. LLE (Locally Linear Embedding) is applied to generate features in lower 9-dimension space. Test images are also transformed into column vectors, embedded into lower dimension, and the corresponding distances are calculated between training and testing images based on features.
Results: LLE is able to preserve the spatial relations of 4DCT from different phases to the lower dimension. LLE-based on features showed high correlation (0.9, p-value 10-4in SI Direction) with target position. We also investigated location of target on the second 4DCT with LLE-based features. We achieved best result when the number of neighbors in LLE was 2, and the number of selected features was 2. There were approximately 2 position deviations (0.7cm) between prediction and the actual positions.
Conclusion: LLE can estimate the location of target inside body with simple skin surface imaging during treatment. This could be used with surface tracking methods such as Align RT and C-Rad to reduce tracking uncertainties.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email