Click here to

Session: Multi-Disciplinary: Deformable Image Registration [Return to Session]

Automatic Liver Tumor Localization Using a Combined Deep Learning and Biomechanical Model (DL-Bio)

H Shao1*, X Huang2, M Folkert3, J Wang4, Y Zhang5, (1) The University of Texas Southwestern Medical Center, Dallas, TX, (2) Northwestern memorial hospital, Chicago, IL, (3) The University of Texas Southwestern Medical Center, Dallas, TX, (4) UT Southwestern Medical Center, Dallas, TX, (5) UT Southwestern Medical Ctr at Dallas, Dallas, TX


MO-IePD-TRACK 4-1 (Monday, 7/26/2021) 5:30 PM - 6:00 PM [Eastern Time (GMT-4)]

Purpose: Liver tumor localization is challenging for X-ray-based image-guided radiotherapy, due to the low tumor contrast against neighboring normal tissues. We developed a liver tumor localization technique through combining 2D-3D deformation, deep-learning, and biomechanical-modeling (DL-Bio) to automatically localize liver tumors through solved deformation-vector-fields (DVFs).

Methods: DL-Bio solves DVFs between planning CT images and new on-board cone-beam projections for tumor contour propagation and localization. It initializes the DVFs through 2D-3D deformation, which solves DVFs by intensity-matching digitally-reconstructed-radiographs of the deformed prior image to on-board cone-beam projections. DL-Bio further applies a deep-learning model to fine-tune the 2D-3D DVFs at the liver boundary, especially for the caudal liver side where lacks sufficient intensity contrast for 2D-3D deformation. The deep-learning model corrects the DVFs at the caudal liver boundary, using a liver surface ring structure, 2D-3D DVFs at the high-contrast cranial liver boundary, and liver boundary motion correlations learnt through model training. Final intra-liver DVFs, including those around the tumor region, were derived with a biomechanical model using the deep-learning-optimized liver boundary DVFs as the boundary condition for finite element analysis.We trained and validated the deep-learning model of DL-Bio using 24 liver patients, and evaluated the combined DL-Bio framework using an independent set of 10 patients. The tumor localization accuracy of DL-Bio was evaluated using the tumor DICE coefficient and center-of-mass-error (COME).

Results: On a total of 90 evaluation images and tumor contours, the average(±s.d.) liver tumor COMEs of the 2D-3D deformation (2D-3D), 2D-3D deformation with biomechanical-modeling (2D-3D-Bio), and DL-Bio were 4.7±1.9 mm, 2.9±1.0 mm, and 1.7±0.35 mm, respectively. The corresponding average(±s.d.) DICE coefficients were 0.60±0.120, 0.71±0.072, and 0.78±0.028, respectively.

Conclusion: The deep-learning model of DL-Bio successfully captures the motion correlations at the liver boundaries for DVF correction, and generates more accurate boundary conditions for biomechanical modeling to improve liver tumor localization accuracy.

Funding Support, Disclosures, and Conflict of Interest: The study was supported by funding from the National Institutes of Health (R01CA240808) and from the University of Texas Southwestern Medical Center.



    Deformation, Localization, Cone-beam CT


    IM/TH- Image Registration Techniques: Machine Learning

    Contact Email