ESTRO 38 Abstract book

S1050 ESTRO 38

Conclusion These results suggest that radiomics can help in detecting, under specific circumstances (e.g. Skewness greater or equal than 0.5) local recurrence at 12 months after SABR and that this decision support system could potentially allow for early salvage therapy. A multicentric study in order to increase the number of patients and to confirm the interesting results is ongoing. EP-1929 Prediction of voxelwise mandibular osteoradionecrosis maps in HNC patients using deep learning L. Humbert-Vidan 1 , I. Oksuz 2 , V. Patel 3 , A.P. King 2 , T. Guerrero-Urbano 4 1 Guy's and St Thomas' NHS Foundation Trust / King's College London, Radiotherapy Physics, London, United Kingdom ; 2 King's College London, School of Biomedical Engineering \& Imaging Sciences, London, United Kingdom ; 3 Guy's and St Thomas' NHS Foundation Trust, Oral Surgery, London, United Kingdom ; 4 Guy's and St Thomas' NHS Foundation Trust, Clinical Oncology and Radiotherapy, London, United Kingdom Purpose or Objective Head and neck cancer (HNC) incidence accounts for 3% of all cancers in the UK. Radiotherapy is one of the main treatments for HNC, either alone or combined with chemotherapy, surgery or hormones. Radiation-induced toxicity to healthy tissue can be a limiting factor for the successful treatment of HNC. Mandibular osteoradionecrosis (ORN) is one of the most severe complications in patients with HNC undergoing radiation therapy. Normally, the risk of toxicities such as ORN is assessed using dose-volume histograms (DVH), but DVH- based parameters ignore the spatial component of radiation dose distribution. We propose the use of deep learning to predict 3D ORN toxicity maps based on 3D dose distribution maps and CT volumes using a U-net convolutional neural network. The ability to predict spatial toxicity maps has the potential to assist decision making in the radiotherapy treatment planning process. Material and Methods Our current deep CNN design is based upon the original U- net deep CNN (see Figure 1). Our model is trained to predict 2D ORN toxicity maps based on corresponding slices extracted from 3D dose distribution maps and CT volumes. The CT and dose map imaging data were used as two channels of input to the network. A total of 232, 59 and 16 sets of slices (i.e. dose map, CT slices and ORN masks) were used for training, validation and testing respectively. Data augmentation was implemented on the training dataset. The Dice Similarity Coefficient loss function was used to train and validate the CNN and the Adam optimiser was used to minimise the loss function. Dropout regularisation (20%) was used to prevent the model from overfitting. A basic manual grid search of parameter values was performed to find the optimal combination of training parameters for the CNN. Based on this search, the best performance was achieved by training the network for 500 epochs with a batch size of 40 and a learning rate of 0.00001.

Results The training, validation and testing Dice coefficients obtained so far are 0.90 (max), 0.51 (max) and 0.13 (mean), respectively. An improvement in all three coefficients was observed when training the network with the augmented dataset. A maximum test dice coefficient of 0.83 was obtained (see Figure 1). However, in the current ORN predictions of the network, some segmentations appear on the wrong side of the mandible; adding more training cases could help to improve the prediction of ORN location. In addition, some predictions do not contain any ORN at all. A custom loss function could be considered to try to improve performance in such cases.

Conclusion Our current design of the U-net is able to produce ORN masks based on CT images and dose distribution maps. The results obtained so far are encouraging but the accuracy of the ORN predictions could be improved. Further work is required in both data processing, CNN design and parameter optimisation to achieve these improvements. The predicted toxicity spatial maps will provide more specific information on what normal tissue regions are at a higher risk. EP-1930 Hypoxia induced by vascular damage could impact on the outcome of stereotactic body radiotherapy E. Lindblom 1 , A. Dasu 2 , I. Toma-Dasu 3 1 Stockholm University, Medical Radiation Physics, Stockholm, Sweden ; 2 The Skandion Clinic, The Skandion Clinic, Uppsala, Sweden ; 3 Stockholm University and Karolinska Institutet, Department of Physics and Department of Oncology and Pathology, Stockhoml, Sweden