Comprehensive Summary
This paper proposes a new, deep-learning based method, DE-subtraction (DES), to synthesize DE fluoroscopy for markless tumor tracking (MTT). A residual U-net model, trained from a digital phantom, was used to synthesize DES from single-energy (SE) x-ray images. This allows for the suppression of bone structures and enhancement of soft tissue, improving tumor visibility. To assess the clinical compatibility of the model for MTT, 10 cases of x-ray fluoroscopic image sequences were acquired from lung cancer patients receiving stereotactic body radiation therapy. The resulting synthesized DES images successfully eliminated rib bones, though spinal bones remained visible, and in one case (case 10) soft tissue texture was partly lost, reducing tumor edge clarity. The model demonstrates improved tumor tracking ability, with the average tracking error (RMSE) being reduced from 1.80 to 1.68 mm and the tracking success rate (TSR) increasing from 50.2 to 54.9%. These findings suggest deep learning-based DES synthesis as a promising approach for MTT, offering an alternative to hardware-dependent DE imaging in lung cancer treatment.
Outcomes and Implications
This method holds significant clinical implications for the advancement of tumor visibility in real-time, markerless tumor tracking for lung cancer radiation therapy. The DES model, under comparison to the raw SE x-ray images, demonstrates potential to significantly reduce errors and improve tracking success rate for some patients, with clinical case 5 seeing a 27% reduction in RMSE and 85.8 to 98.9% improvement in TSR. The proposed method also demonstrates real-time capability with a total processing time of 55 ms per frame, which is within the 66.7 ms per frame interval of current standard kV x-ray imaging capturing. The sequence is automated after the first frame and has minimal sensitivity to minor misalignments during affine registration, increasing its usability in clinical settings.