Automatic Cone Beam Projection-based Liver Tumor Localization by Deep Learning and Biomechanical Modeling

Abstract

Accurate on-board liver tumor localization remains a challenge for CBCT-based image-guided radiotherapy, due to tumors’ low-visibility within normal liver tissues. Deformable registration-driven contour propagation from planning CTs may facilitate automatic, accurate tumor localization. We developed a technique combining a population-based liver surface deformation model and a patient-specific, biomechanical modeling-guided algorithm to achieve accurate deformation-driven liver tumor contour propagation and localization. The technique solves the deformation-vector-fields (DVFs) for tumor propagation by three steps: 1. Solving a coarse DVF via 2D-3D deformable registration by intensity-matching DRRs of the deformed planning CT to on-board cone-beam projections; 2. Feeding the DVF into a population-based deep learning model to improve its accuracy around liver surface region; and 3. Applying the accuracy-enhanced liver surface DVF to a patient-specific liver biomechanical model as the boundary condition, to solve intra-liver DVF through finite element analysis. The optimized intra-liver DVF is then applied to the liver tumor contour to propagate it onto the CBCT frame for automatic localization. For step 2, the deep learning model uses a U-net architecture to perform voxel-wise correction to the 2D-3D DVF solved by step 1, especially at the caudal liver surface region where limited tissue contrast diminishes the accuracy of 2D-3D registration. The population-based model was trained on a total of 21 patients through supervised learning, using patient-specific 2D-3D DVF and liver contour pairs as input, and high-accuracy liver surface DVF as ‘gold-standard’ output. The ‘gold-standard’ DVF was obtained by manually contouring, density-overriding, and deform-registering liver organs on corresponding CT-CBCT pairs. We evaluated the technique on 10 liver patients not included in the training list. Each patient had a contrast-enhanced 4D-CT set, of which the 0% phase was used as the planning CT, and phases 10-90% as 4D-CBCTs to simulate 4D cone-beam projections for tumor localization evaluation. Liver tumors were contoured by a radiation oncologist on all 4D-CT phases for propagation (phase 0%) or reference (phases 10-90%). Both DICE similarity index and center-of-mass-error (COME) were calculated between propagated liver tumor contours and physician’s reference contours for quantitative analysis. Using 20 cone-beam projections for each localization, the 2D-3D registration technique, biomechanical modeling technique (without deep learning), and our technique localized the liver tumors to average ± s.d. COMEs of 4.5 ± 1.4 mm, 2.5 ± 0.8 mm, and 1.6 ± 0.5 mm, and DICEs of 0.61 ± 0.23, 0.73 ± 0.16, and 0.80 ± 0.09, respectively. The developed technique enables automatic, accurate liver tumor localization (<2 mm) using few projections (< = 20), which facilitates 4D tumor localization after projection phase sorting.