Room: Track 2
Manual on-board liver tumor localization is challenging due to the limited tumor contrast in CBCT images. Automatic tumor localization by propagating tumor contours from planning CT to CBCT using deformation-vector-fields(DVFs) can be efficient and accurate. We developed a DVF-driven liver tumor localization technique by combining 2D-3D deformable registration, deep learning and biomechanical modeling(U-net-Bio).
U-net-Bio started with 2D-3D deformable registration to derive a coarse DVF through intensity-matching DRRs of the deformed CT to on-board cone-beam projections. The outcome 2D-3D DVF was subsequently fed into a deep learning framework(U-net), to allow voxel-wise DVF correction to improve its accuracy at the liver boundary, especially the limited-contrast lower liver boundary. Lastly, U-net-Bio feeds the liver boundary DVF into biomechanical modeling to derive intra-liver DVF to propagate CT tumor contours onto CBCT for localization.
The U-net model was derived via supervised learning on 25 liver 4D-CT sets. From each 4D-CT, we viewed phase 0% as planning CT, and used simulated cone-beam projections from other phases to derive 2D-3D DVFs. The 2D-3D DVFs and the liver contour at phase 0% were used as U-net input. The high-quality output DVFs for training were derived by directly registering the CT-CT pairs using Demons algorithm, with livers on both CTs density-overrode for boundary enhancement.
We used 4D-CT sets of another 7 patients to evaluate U-net-Bio. Similarly, phase 0% was used as planning CT, and other phases(10%-90%) were used to simulate projections for tumor localization. The 4D-CTs were contrast-enhanced and the tumors were manually contoured for propagation(phase 0%) and localization accuracy evaluation(phases 10%-90%).
Using 20 projections, U-net-Bio localized liver tumors to an average center-of-mass-error of 1.8±0.5mm and DICE coefficient of 0.79±0.09, compared to 4.3±1.4mm and 0.61±0.23 for 2D-3D deformation.
U-net-Bio localizes liver tumors to within 2 mm using only 20 projections, allowing efficient and accurate IGRT.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by grants from the US National Institutes of Health (R01 CA240808 and R01 EB027898), and from a UT Southwestern Radiation Oncology Departmental Seed Grant.