Click here to

Session: Adaptive Particle Therapy [Return to Session]

Deep Learning Based 4D Synthetic CTs Generated From CBCTs for Proton Dose Calculations in Adaptive Proton Therapy

A Thummerer1*, C Seller Oria1, S Visser1, P Zaffino2, A Meijers3, R Wijsman1, G Guterres Marmitt1, J Seco4,5, J Langendijk1, A Knopf1,6, M Spadea2, S Both1, (1) Department of Radiation Oncology, University Medical Center Groningen, University of Groningen, Groningen,NL, (2) Department of Experimental and Clinical Medicine, Magna Graecia University, Catanzaro,IT, (3) Center for Proton Therapy, Paul Scherrer Institute, Villigen, CH, (4) DKFZ Heidelberg, Heidelberg, DE, (5) Department of Physics and Astronomy, Heidelberg University, Heidelberg, DE (6) Department I of Internal Medicine, Center for Integrated Oncology Cologne, University Hospital of Cologne, Cologne, DE

Presentations

SU-H300-IePD-F5-3 (Sunday, 7/10/2022) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 5

Purpose: To generate deep learning based 4D synthetic CTs (sCTs) from sparse-view 4DCBCTs and evaluate their suitability for proton dose calculations in adaptive proton therapy workflows. Additionally, 4D-sCTs were also compared to deep learning based 3D-sCTs.

Methods: A U-net like deep convolutional neural network was trained to generate 4D-sCTs, using a dataset containing 4D-CTs and same-day 4D-CBCTs of lung cancer patients treated with proton therapy (28 training, 4 validation, 18 testing). For training and evaluation, the 4D-CTs were deformed to the 4D-CBCTs (phase-by-phase). Image similarity between sCTs and CTs was assessed via mean absolute error (MAE). Treatment plans were recalculated on the sCTs and the resulting dose distributions were compared globally (3%/3mm gamma analysis) and locally (CTV, lung, heart, esophagus) against dose distributions calculated on 4DCTs. Evaluation was performed for maximum inhale and exhale 4D-sCT phases (0% and 50%), averaged 4D-sCT and a 3D-sCT.

Results: Individual phase images resulted in an average MAE of 47.5±6.1 HU and 48.1±6.5 HU for the 0%- (max. inhale) and 50%-phase (max. exhale) respectively. For the 4D-sCT average, a MAE of 37.7±6.2 HU was observed, which was comparable to the 3D-sCT MAE of 37.7±7.7 HU. The gamma analysis resulted in average pass ratios of 93.3±4.8%, 92.7±4.9%, 94.3±4.7% and 92.6±5.15%, for 0%-phase, 50%-phase, 4D-sCT average and 3D-sCT respectively. Dose differences for the CTV (D98) were within ±1.7% for all patients and sCTs. Larger mean dose differences were observed in organs-at-risk: up to 7.7% in the heart (3DsCT), up to 4.9% in the esophagus (4DsCT 0%) and up to 2.2% in the lung (4DsCT 50%).

Conclusion: 4D-sCTs showed similar dose calculation accuracy as 3D-sCTs, despite the lower image quality found in individual phases. 4D-sCTs could offer the possibility to account for breathing motion while preserving dose calculation accuracy in adaptive proton therapy workflows.

Funding Support, Disclosures, and Conflict of Interest: Langendijk JA is a consultant for proton therapy equipment provider IBA. University of Groningen, University Medical Centre Groningen, Department of Radiation Oncology has active research agreements with RaySearch, Philips, IBA, Mirada, Orfit. This study was financially supported by a grant from the Dutch Cancer Society (KWF research project 11518)

Keywords

Cone-beam CT, Image Processing, Protons

Taxonomy

IM/TH- Cone Beam CT: Machine learning, computer vision

Contact Email

Share: