Prediction of mandibular ORN with DL-based classification of 3D radiation dose distribution maps
PO-1770
Abstract
Prediction of mandibular ORN with DL-based classification of 3D radiation dose distribution maps
Authors: Laia Humbert-Vidan1,2, Vinod Patel3, Robin Andlauer2, Andrew P King2, Teresa Guerrero Urbano4,5
1Guy's and St Thomas' NHS Foundation Trust , Radiotherapy Physics , London, United Kingdom; 2King's College London, School of Biomedical Engineering & Imaging Sciences, London, United Kingdom; 3Guy's and St Thomas' NHS Foundation Trust, Oral Surgery, London, United Kingdom; 4Guy's and St Thomas' NHS Foundation Trust, Clinical Oncology, London, United Kingdom; 5King's College London, Cancer Clinical Academic Group, London, United Kingdom
Show Affiliations
Hide Affiliations
Purpose or Objective
Absorbed radiation dose to the mandible is an important risk factor in the development of mandibular ORN in patients with HNC. Dosimetric parameters typically included in ORN risk prediction studies are DVH-based, and localisation information is thus lacking. We propose the use of a binary classification 3D convolutional neural network (CNN) to extract the relevant dosimetric information from mandibular 3D radiation dose distribution maps and predict the incidence of ORN.
Material and Methods
A total of 70 ORN cases were retrospectively selected from our clinical database and matched with 70 control cases based on primary tumour site. For the purpose of this study, any grade of ORN was considered an event. The mandibular radiation dose distribution volume was calculated based on the RT plan and mandible structure DICOM files. The 3D DenseNet CNN was implemented for binary classification using the MONAI Pytorch-based framework. A Softmax activation layer was added at the end to obtain the predicted probability for each class. The network was trained on 3D dose distribution volumes for 200 epochs with a dropout rate of 0.8 and a step-decay learning rate schedule of step size 50 and gamma factor 0.5. A class-balanced test dataset was kept aside for inference. Stratified cross validation was applied to the rest of the data for hyper-parameter optimisation. Small 3D random rotation and zoom augmentations were applied to the training images with a batch size of 5. A soft voting ensemble was created with 9 repeats of the final CNN to compensate for model variance in the test class label prediction accuracy results.
A hard voting ensemble of 9 random forest (RF) binary classifiers was implemented with mandible DVH-based metrics for comparison to the CNN-based results.
Results
The 3D CNN was able to predict ORN incidence on the independent test dataset with an ensemble accuracy of 75% (range 71% to 86%). The ensemble accuracy obtained with the RF binary classifier was also 75% (range 75% to 78%).
Conclusion
Our results support the use of 3D dose distribution maps as an alternative to DVH-based dosimetric variables in ORN prediction models. Clinical decisions such as pre-RT dental extractions or radiation dose reduction would benefit from knowledge about which mandible region is more likely to develop ORN for a particular patient. This study is a first step towards individualised prediction of mandibular ORN localisation probability.