3167 - Deep Learning-Based Anatomical Recovery for Dose-Accurate MR-Only Radiotherapy Planning

Purpose/Objective(s): In MR-only radiotherapy planning (MROP), limited field-of-view (LFOV) acquisition and imaging artifacts can introduce truncated anatomy and density inaccuracies, reducing the reliability of dose calculation and preventing comprehensive plan evaluation due to incomplete visualization of organs at risk (OAR). These limitations are particularly pronounced in Vestibular Schwannoma (VS) patients, where high-resolution MR imaging is typically focused on the internal auditory canal and cerebellopontine angle, resulting in incomplete visualization of surrounding skull-base and neck anatomy required for accurate dose computation. The purpose of this study was to develop and evaluate a data augmentation–driven deep learning (DL) framework to recover planning-appropriate images from truncated or corrupted inputs, enabling dose-accurate MROP in clinically relevant LFOV scenarios.

Materials/Methods: The proposed framework consists of two sequential DL models. First, a conventional MR-to-synthetic CT (MR2sCT) model was used to convert truncated MR images into truncated synthetic CT (sCT), providing voxel-wise electron density information for dose calculation while preserving the original MR acquisition constraints. Second, an image recovery network was applied to address anatomical truncation. This recovery model (sCT2sCTx) was trained to restore missing anatomical regions beyond the LFOV by learning from full-length reference CT (rCT) volumes. During training, truncation was synthetically introduced using a data augmentation–driven strategy, in which longitudinal box masks with randomly sampled sizes and spatial locations were applied on-the-fly to simulate clinically observed LFOV patterns, particularly in skull-base imaging. During inference, the recovery network generated extended sCT (sCTx) volumes from truncated sCT inputs, yielding anatomically completed images suitable for comprehensive dose calculation and OAR evaluation. The sCT2sCTx model was implemented as an auto-encoder adapted from nnU-Net and trained using full-length rCT as ground truth.

Results: A brain case was evaluated by the proposed framework. The sCTx achieved reliable anatomical compensation and tissue contrast compared with the rCT, including truncated regions. Dose calculation performed on sCTx demonstrated high agreement with that based on the rCT, as reflected by consistent DVHs and a 98.67% gamma passing rate using 3%/3 mm criteria between the dose distributions.

Conclusion: A data augmentation–driven DL framework was developed to recover truncated anatomy in LFOV medical images. In MROP applications, the recovered images supported accurate dose calculation and more complete plan evaluation by restoring missing anatomical regions, enhancing robustness in the presence of LFOV constraints.