ESTRO 2024 - Abstract Book

S3166

Physics - Autosegmentation

ESTRO 2024

resource-limited settings. This pilot study aims to evaluate the feasibility of using a machine learning-based segmentation model to accurately delineate the hippocampus using only planning CT scans. The development and validation of such a model have the potential to expedite patient treatment by eliminating delays related to MRI scheduling.

Material/Methods:

A total of 76 patients undergoing either HA-WBRT or HA-PCI in years 2020-2022 at our facility were included in the study. All subjects underwent high-resolution MRI prior to treatment, and these images were rigidly co-registered with the corresponding planning CT scans. Hippocampal contours were generated in accordance with the Hippocampal Sparing NRG Atlas for RTOG 0933. A deep learning-based segmentation model was developed using nnU-Net v2. Five-fold cross-validation was used for internal validation and the model's performance was evaluated using average Dice coefficient for all folds.

Results:

Our machine learning algorithm achieved an average cross-validation Dice coefficient of 0.721, indicating a relatively high degree of overlap between the model's segmentation and the gold standard contours. To facilitate further research and model development, we have made the anonymized dataset, model files, and source code publicly accessible at GitHub Repository - https://github.com/kstawiski/hippohelper. This repository can serve as a foundational resource for training and validation in future multi-institutional studies aimed at enhancing the robustness and clinical applicability of CT-based hippocampal segmentation models.

Conclusion:

Our initial findings substantiate the viability of CT-based hippocampal delineation. Nevertheless, the compilation of a more expansive, multi-institutional dataset is imperative for the development and validation of a clinically deployable model.

Keywords: Hippocampal avoidance, automatic segmentation