ESTRO 2024 - Abstract Book

S5029

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

In order to tackle the severe class imbalance during training, a balanced class sampling approach was adopted using TorchIO [4] such that 50% of the patches within the batch will have central voxels with radionecrosis. The model was trained using cross-entropy loss and was optimized for pseudo-receiver operating characteristic-area under the curve (ROC-AUC) estimated on class-balanced patches sampled from the validation set. During inference, the predictions were made on the entire input images based on a sliding window approach to estimate the actual roc-auc amongst other performance metrics. The window size was set to be the patch size used during training while the stride was configured such that the adjacent predictions overlap by 50%. Stratified 5-fold cross-validation was adopted such that an equal percentage of patients with/without necrosis were present across each fold to estimate a robust predictive performance of the model.

Results:

Retrospective radiological and dosimetric data of 41 patients, including 36 (88%) meningiomas and 5 (12%) hemangiopericytomas, were evaluated for this study. The median prescribed dose was 60 Gy (range: 54-74 Gy). 26 (63.4%) patients reported brain radionecrosis after a median time of 17.5 months (range: 4.6-53.9 months), at the end of PBT. The model demonstrated a mean voxel-wise cross-validation ROC-AUC score of 0.9072 (std = 0.0988) with a Brier score of 0.0014 (std=0.0005). The mean balanced accuracy was also found to be 0.8633 (std = 0.1105).

Conclusion:

Although this study requires rigorous validation, a deep learning-based multi-modal data analysis holds the potential to be used for early assessment of BRN, to facilitate PBT optimization in meningioma and hemangiopericytoma patients.

Keywords: deep learning, prediction, radionecrosis