Auto-segmentation models used for manuscript: "NTCP models predicting acute and late bowel toxicity after total neoadjuvant treatment or chemoradiation for locally advanced rectal cancer in the RAPIDO trial"

Zenodo (CERN European Organization for Nuclear Research)

Abstract

For this manuscript, two auto-segmentation models were trained in-house to predict the bowel contours according to the EMBRACE and RTOG delineation guidelines. For both models, we used a model from TotalSegmentator (v1.5.7)1 to serve as a base model and iteratively fine-tuned the model using nnUNet (v1.7.0)2 to improve the contouring performance. In total six rounds of fine-tuning were needed until only minimal corrections per prediction were required. For the first round of fine-tuning, an external dataset from the Leiden University Medical Center (LUMC) was used (train: 58 cases, validation: 15 cases). As input for the subsequent fine-tuning rounds, we used the model of the previous round to predict cases of the RAPIDO dataset and carefully corrected the predictions. This resulted in the following fine-tuning rounds with N cases for training and validation: round 2: (train: 7, val: 2), round 3 (train: 11, val: 3), round 4 (train: 15, val: 4), round 5 (train: 24, val: 6), round 6 (train 37: val: 10). In addition, two extra auto-segmentation models were trained for patients in prone position by fine-tuning the models from round 6 with a dataset of patients in prone position (train: 37, val: 10). All predicted bowel contours were manually reviewed and corrected if necessary. Models: - Task768_EMBRACE_BB: Model predicting the bowelbag according to the EMBRACE delineation guidelines. - Task769_RTOG: Model predicting the bowelbag according to the RTOG delineation guidelines. - Task770_EMBRACE_BB_prone: Model predicting the bowelbag according to the EMBRACE delineation guidelines for CT scans acquired in prone position. - Task771_RTOG_prone: Model predicting the bowelbag according to the RTOG delineation guidelines for CT scans acquired in prone position. Installation: Use the nnUNet (v1.7.0) command “nnUNet_install_pretrained_model_from_zip” to install the models in your nnUNet trained models folder. Inference: The models use a single CT as input. For inference run the following command: “nnUNet_predict -i {data_dir} -o {output_dir} -t {task_model} -tr nnUNetTrainerV2_ep4000_nomirror -p nnUNetPlans_pretrained_FINETUNE_BOWELS -m 3d_fullres”

Authors 0

1. Author list not loaded yet.