Room: AAPM ePoster Library
Tracking the tumor in the beam’s-eye-view is challenging for VMAT treatments due to (1) partial occlusions of the target by the MLC leafs and (2) the low contrast of the tumors on EPID images. We introduce a novel markerless beam’s-eye-view tracking based on semantic deep learning segmentation.
The tracking system consists of the trained U-Net for initial segmentation and an image registration component. The registration algorithm combines the current aperture, identified using a level-set method, with the projected tumor contour to refine the segmentation.
EPID images from 9 patients (2 fractions/patient) treated with VMAT for non-small cell lung cancer were manually labelled using projected CT tumor contours. Images from 3 patients were used to train a U-Net (VGG19) architecture for semantic segmentations.
The system was tested on EPID images from the remaining 12 fractions from the 6 unseen patients.
We evaluated the error in tumor centroid motion in the x-(lateral) and y-(superior/inferior) directions of the beam’s-eye-view images by comparing with manual labelled ground-truth. Dice Similarity Coefficients (DSC) between the segmented tumors and the ground-truth labels were computed.
The tracking system has a geometric error of 0.2±4.5 mm in the x-direction and 0.3±5.5 mm in the y-direction of the beam’s-eye-view images on the test dataset. The DSC of the segmented tumor on patients’ beam’s-eye-view images compared to the ground-truth was 0.69±0.18 (N=23160 images) with a failure rate (DSC=0) of <0.5%.
A direct tumor tracking system for beam’s-eye-view images of lung cancer patients treated with modern VMAT plans was developed and evaluated. The tracking system has good agreement with the tumour shape, overcoming challenges of beam’s-eye-view tracking. This work represents the first step toward reaching for the ideal real-time tumor tracking, one that requires no implanted markers adds no imaging dose and operates without any additional hardware.
Funding Support, Disclosures, and Conflict of Interest: D T Nguyen is funded by Australian NHMRC and Cancer Institute NSW Early Career Fellowships. P Keall acknowledges funding from NHMRC Fellowship.