Prior-Guided Automatic Segmentation of Brain Tumor Targets for Adaptive Radiation Therapy on MR-Linacs
Abstract
Purpose
T1-weighted (T1w) and T2-weighted (T2w) FLAIR MRI provide complementary image contrast for delineating gross tumor volume (GTV) and clinical target volume (CTV) in brain tumor radiotherapy (RT) planning during both MRI simulation (MR-Sim) and MRI-guided RT (MRIgRT). In MR-LINAC-based adaptive RT (ART), on-board MRI is acquired at each fraction to update GTV/CTV contours prior to plan re-optimization. However, manual re-contouring across multiple MRI sequences is time-consuming for clinicians (typically 20-30 minutes), substantially prolonging treatment procedure sessions on MR-LINACs compared with conventional workflows. In this study, we propose an automatic tumor segmentation approach that integrates multiple-sequence MRI with prior information from MR-Sim and previous treatment fractions to enable efficient and consistent target delineation for adaptive MRIgRT.
Methods
We retrospectively collected MR-LINAC-based ART data from 75 brain tumor patients. Each patient underwent 3-6 treatment fractions, with most fraction including T1w and T2w-FLAIR MRI and multiple target contours. The dataset was divided into training, validation, and test sets. We developed a prior-guided, multi-modality LSTM-UNet model that use imaging information from multiple MRI sequences and ground-truth target contours drawn in previous fractions. The LSTM units capture temporal information across treatment fractions, enabling the model to leverage prior anatomical and tumor evolution knowledge to predict target contours for the current fraction.
Results
We evaluated the proposed multi-modality LSTM-UNet against a baseline UNet and a traditional image-registration-based method. The multi-modality LSTM-UNet achieved the highest segmentation performance, with an average Dice similarity score of 84.5%, compared with 82.2% for the image-registration method and 83.0% for the baseline UNet.
Conclusion
By integrating multi-sequence MRI and prior knowledge from previous treatment fractions, the proposed method improves segmentation accuracy for brain tumor targets in MRI-guided ART on the MR-LINACs. Compared with traditional single-image-based models, leveraging complementary imaging contrasts and temporal prior information is essential for accurate target delineation.