Click here to

Session: Automation in Treatment Planning [Return to Session]

Automating Contouring and Treatment Planning for Pediatric 3D-Craniospinal Irradiation Therapy

S Hernandez1,2*, J Parkes3, H Burger3, C Nguyen2, D Rhee2, A Paulino2, T Netherton1,2, R Mumme2, J Guma-de La Vega2, J Duryea2, C Cardenas4, R Howell1,2, D Fuentes1,2, J Pollard-Larkin1,2, L Court1,2, (1) UT MD Anderson Cancer Center UTHealth Graduate School of Biomedical Sciences, Houston, TX (2) The University of Texas MD Anderson Cancer Center, Houston, TX (3)Department of Radiation Oncology, Groote Schuur Hospital and University of Cape Town, Cape Town, South Africa (SA) (4) Department of Radiation Oncology, University of Alabama at Birmingham, Birmingham, AL


MO-H345-IePD-F7-1 (Monday, 7/11/2022) 3:45 PM - 4:15 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 7

Purpose: Pediatric medulloblastoma patients in LMICs are treated with photon craniospinal irradiation(CSI). We partnered with international clinicians to develop and test a (1)auto-contouring and (2)CSI auto-planning tool robust to pediatric CSI treatment field variations, specifically for resource-constrained settings.

Methods: (1)To contour 13 planning structures, a nn-UNet model was trained:tested(80:20 ratio) on 164 pediatric medulloblastoma patient CT scans(aged:2-19 years, median=7) with five-fold cross-validation during training. (2)Two lateral brain fields are automatically matched to (a)single spine field(100cm SSD), (b)extended single spine field(120cm SSD), or (c)two spine fields(100cm SSD). Field configurations are automatically selected via auto-calculated spine length. The brain isocenter is automatically placed between the mandible and shoulder auto-contours. 1-2 spine isocenters are automatically placed (depending on calculated spine field size). Lateral brain fields conform to the brain target auto-contour(brain+cribriform-plate+upper-canal auto-contours). Posterior spine fields automatically conform to the vertebral column auto-contour. Feathering was implemented at each match line(yielding 9-12 fields). Auto-contours are employed for QA checks (i.e.mandible to assess field divergence). The auto-planning tool was tested on 7 patients(ages:3-10 years) to generate auto-contours and CSI plans(3 field configurations). Clinicians reviewed the outputs of each step of the integrated auto-planning tool. Target coverage and dose to auto-contours was quantified.

Results: The average DSC and HD(cm) for auto-contours ranged from 0.73-0.99(8/13 structures>0.80) and 0.09-3.10cm(8/13 structures<1cm), respectively. The algorithm correctly placed isocenters, generated fields, optimized field match lines, and feathered junctions for all patients. The average V95(brain/spinal canal) for single, extended, and multi-field configurations was 99.9±0.06%/99.5±0.83%, 99.7±0.07%/99.6±0.90%, 99.5±0.07%/99.5±0.90%, respectively(Rx=23.4Gy/13fx). V95(cribriform-plate) was 96.2±2.60%. V20(kidneys) was 5.54±3.07%. Mean dose to the eyes, lenses, thyroid, brainstem, heart and lungs were 15.8±1.97, 6.00±1.97, 17.0±2.35, 23.7±0.08, 11.3±1.10, and 2.56±0.15Gy, respectively.

Conclusion: The auto-planning tool successfully generated pediatric normal tissue contours for varying patient sizes and CSI plans for multiple field configurations and may be a useful tool for resource-constrained settings.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by the John and Charlene Kopchick research fellowship and the Cancer Prevention Research Institute of Texas (CPRIT) Training Award (RP210028).


Conformal Radiotherapy, CT, DICOM-RT


TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation

Contact Email