Automatic, Real-Time Tumor Tracking on Kv-CBCT Projections Utilizing a Hybrid, AI/Drr-Based Background Subtraction Technique
Abstract
Purpose
For thoracic cancers, respiratory-induced tumor motion must be accounted for to ensure accurate treatment. While real-time tumor tracking represents the ideal in motion management, existing commercial options face significant hurdles to implementation. This study introduces a pipeline for real-time tumor tracking verification using kV-CBCT projection imaging on conventional linacs via a novel, hybrid AI/DRR background subtraction technique.
Methods
CBCT volumes and raw kV projections were prospectively collected for 57 TrueBeam fractions across 14 Stage I-III non-small cell lung cancer patients. GTV voxel intensities within each CBCT were masked to -1000 HU to generate tumorless-DRRs (tDRRs) and rigidly registered to the corresponding kV projection. Tumors that were visible on kV projections were manually delineated to generate ground truth segmentations (n=15). The remaining 42 fractions were used to train a U-Net for tDRR-to-kV mapping to generate simulated tumorless-kVs (sim-kVs) using noise contrastive estimation and L2 loss for 1000 epochs. Sim-kVs were then subtracted from the raw kVs to suppress surrounding anatomy via background subtraction. Prior to evaluation, the U-Net was fine-tuned over continuous 30° arcs using L1 loss. All losses were calculated on a patchwise level outside the ITV to avoid tumor sampling. Binary thresholding was used to assess contrast enhancement and benchmarked against tDRR subtraction and adaptive histogram equalization (AHE) using Dice Similarity Coefficients (DSC) and L1 centroid errors. Significance was tested using paired t-tests.
Results
Model fine-tuning took 2.28+/-0.48 min per arc on a NVIDIA Quadro RTX 4000. Our method significantly outperformed other methods (p<0.001), with median DSC and centroid errors for sim-kV, tDRR, and AHE methods at 0.81, 0.21, 0.05 and 1.58, 13.8, 15.41 mm, respectively.
Conclusion
We demonstrate the feasibility of an onboard markerless, real-time tumor tracking pipeline verification method. Our technique outperforms classical methods and can be implemented in real time with modest computing resources.