A Patient-Specific Deep Learning Model for Angle-Agnostic Tumor Tracking In Projections for Lung Cancer Radiotherapy
Abstract
Purpose
Radiotherapy for lung cancer is challenged by respiratory motion and inter-fractional anatomical variations. The respiratory motion management based on tumor tracking can reduce the discrepancies between planned and absorbed doses. However, current clinical tumor tracking methods in clinical practice rely on patients’ tolerance and specialized hardware. We developed a patient-specific framework based on deep learning for angle-agnostic tumor tracking in projections of CBCT.
Methods
The 4DCT contained multiple fractions of 7 lung cancer patients were included in this study. Inter-fractional variations and respiratory tumor motion were simulated using deformation vector fields derived from the first-fraction 4DCT, combined with localized Gaussian-filtered deformations. The projections were generated from the simulated CT using DeepDRR, with noises and scatters added via CUT. We proposed a tumor tracking model based on a Siamese attention-based U-Net with a weight-sharing encoder, incorporating a muti-modal change fusion module and a convolutional block attention module in the decoder. The performance and robustness of the proposed model were test on subsequent treatment fractions using the Dice similarity coefficient (DSC) and the 95th percentile Hausdorff distance (HD95), and were compared with models reported in previous studies using paired t-tests.
Results
Compared with models reported in previous studies, our model achieved the best DSC of 0.916±0.036 and HD95 of 4.20±2.46 mm. An improved U-Net architecture achieved a DSC of 0.876 ± 0.065 and an HD95 of 6.97 ± 5.75 mm, and the standard U-Net achieved a DSC of 0.739 ± 0.121 and an HD95 of 17.26 ± 9.07 mm.
Conclusion
The proposed model demonstrated great tracking stability and precision. Our deep learning framework has potential to provide a robust and accurate markerless tumor tracking method for common linear accelerators.