ESTRO 2024 - Abstract Book

S4208

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2024

A long-held goal of radiotherapy is to reduce radiation exposure to adjacent healthy tissue. The MRIdian 0.35 T MR linac (Viewray Inc, USA) allows for intra-fraction motion management via a cine MRI-based breath-hold gating approach, which has been shown to reduce healthy tissue dose compared to ITV or mid-ventilation strategies. However, gating results in prolonged treatment times. To avoid prolonging treatments even further or to enable MLC tracking in free-breathing in the future, precise real-time tumor tracking is required. To address this challenge, Segmentation Residual Networks (SegResNets)[1,2] were employed, a class of U-net algorithms that have achieved remarkable success in the field of medical image segmentation while requiring relatively little data for training. Therefore, they are good candidates for developing patient-specific tumor tracking models. We implemented SegResNets based on clinically acquired 4 Hz or 8 Hz sagittal 2D-cine MRIs and compared the predicted tumor contours against manually generated ground truth. With ethics committee approval, about 5600 raw cine MRIs frames without any annotation were collected for 23 patients (abdomen, heart, liver, lung, mediastinum, pancreas, prostate) treated at our hospital and consecutive frames (up to 240) were extracted for each patient. All cine MRIs were resampled to 1x1 mm2 and cropped to 224x224 pixels. The tumor or a surrogate tracking target for all frames was then manually segmented by two different observers with assistance of clinical data to generate the ground truth for model optimization and evaluation. The patients were partitioned into a validation cohort comprising 10 patients and a testing cohort with 13 patients. Models were trained from scratch for each patient on either 4, 8, or 12 manually segmented frames. We trained the models either for 150 or 300 epochs with image augmentation (affine, bias field and Gibbs noise) to avoid overfitting during patient specific training. Subsequently, the trained models were employed to predict the tumor contours in the remaining frames of each patient. In total, 6 models were developed for each patient of the validation cohort based on combination of the number of training frames and epochs. For testing, we only used the combination of training frames and epochs which led to satisfactory accuracy while minimizing time for contouring and model training. Evaluation metrics included the Dice similarity coefficient (DSC), the 50% Hausdorff distance (HD50%), the 95% Hausdorff distance (HD95%), and root mean squared error of the target center-of-mass in both the anterior-posterior (RMSEAP) and superior-inferior (RMSESI) directions. Material/Methods:

Results:

Based on the validation, we selected the model based on 8 frames and 300 epochs for model testing. Training the models with this combination yielded testing tumor segmentation accuracy (Fig. 1) comparable with inter-observer variability (Tab. 1). Without any dedicated optimization, the process of contouring took approximately 4 minutes, model training about 3 minutes. A marginal increase in prediction accuracy could be obtained by expanding the training dataset to 12 frames. However, this would require additional 3.5 minutes for contouring, thus prolonging the pre-treatment workflow. The inference time of the models remained under 8 ms per frame, which makes the approach suitable for real-time tumor tracking.