Transfer Learning-Based Synthetic Computed Tomography (CT) Method For Daily Proton Therapy Dose Validation For Pediatric Patients With Anesthesia

Document Type

Conference Proceeding

Publication Date

12-2025

Publication Title

International Journal of Particle Therapy

Abstract

Background and Aims: The importance of deep learning-based synthetic computed tomography (sCT) from daily cone beam CT (CBCT) has been recognized in assisting proton treatment adaptive decisions. However, all the previously published sCT models were generated from adult patient population datasets. We proposed a method to adapt those adult patient based model to anesthesia pediatric patient to overcome the limited available paired same-day CBCT and QA-CT Methods: We introduced transfer learning based sCT approach by applying an agent algorithm to transfer the adult-model to the pediatric patient population. A adult patient’s CBCT from a similar disease site is selected to provide geometry information. A virtual adult-geometry-based feature map is extracted for the original adult model by applying an agent function A(.) to adjust the geometry and supporting structures for the specific pediatric CBCT. Considering the HU prediction of each pixel as a probability distribution, given the domain as D, feature space as X, the marginal probability distribution as P(X), transfer learning is to apply A(.) to map the probability function fT (XT) from fS (XS): A(XT) -> XS; fT (XT)- > fS (XS)->P(X) where subscript T stands for target domain and source domain (figure1). A total of five anesthesia pediatric patients who had same-day paired CBCT and QA-CT were selected for evaluation. Mean HU of ROI was used to check the sCT quality. 3D gamma index were used to quantitatively assess the dose reconstruction accuracy. Results: Table-1 shows Δ Mean HU. The gamma passing rate, with 98.51±0.90% under 3%/3mm criteria and 90.56±3.22% under 2%/2mm criteria. Figure2 shows DVHs of the initial plan recalculation on the sCT and QA-CT. Conclusions: This study demonstrates for the first time that the transfer learning approach can convert an adult synthetic CT model for a pediatric patient population, which is specifically important for pediatric patients requiring general anesthesia for proton beam therapy.

Volume

17

Issue

Suppl

First Page

101123

Last Page

101123

Comments

63rd Annual PTCOG (Particle Therapy Cooperative Group) Conference, June 2-7, 2025, Buenos Aires, Argentina

DOI

10.1016/j.ijpt.2025.101123

Link to Full Text

Share

COinS