ESTRO 37 Abstract book

ESTRO 37

S477

predict the exact number of necessary seeds.

Figure 1. Measured (Farmer) and calculated (MC and RT), out-of-field profiles for 50mm IRIS collimator and measured 0-field profile (closed IRIS) are presented for. A) SSD 500mm (depth 15mm). B) SSD 785mm (depth 15mm). C) SSD 785mm (depth 300 mm). Conclusion MC based dose calculation algorithm in MultiPlan TPS underestimated the out-of-field doses, and thus, also the mean doses in water phantom. RT algorithm was highly inaccurate in out-of-field region. The head leakage might have a significant contribution to the whole body mean dose. PO-0897 Predicting the number of seeds in LDR prostate brachytherapy using machine learning and 320 patients N. Boussion 1 , A. Valeri 2 , J. Malhaire 3 , D. Visvikis 1 Purpose or Objective The number of iodine-125 seeds used for a given patient in LDR prostate brachytherapy is usually ordered according to the expected prostate volume. However, the real volume may be found different the day of the exam (Fig.1), leading to a high number of remaining radioactive seeds that need to be stored and managed according to legal and strict radiation protection procedures. The aim of the proposed study was to use machine learning on a 320 patients' database in order to build an abacus able to 1 INSERM UMR 1101, LaTIM, Brest, France 2 CHRU BREST, Urology, Brest, France 3 CHRU BREST, Radiotherapy, Brest, France

Material and Methods For each one of our 320 patients, 31 pre- and post- treatment parameters were regularly recorded. These parameters were used as an input in a series of machine learning algorithms for both classification and regression purposes. For classification, patients were separated into two groups according to their remaining seeds (higher or lower than 20 remaining seeds). For regression, the objective was to retrieve the exact number of implanted seeds. Models were trained and validated using 66.6% and 33.3% of the dataset respectively. Ten-fold cross validation was used on the training dataset in order to get statistics and to avoid overfitting. The algorithm which provided the best fitting was then used to build an abacus able to predict the number of seeds for future patients. Results For both classification and regression, the Logistic Regression algorithm was found to provide the best results. Classification accuracy was 96.7%+/-4.2% and 96.2% for training and validation groups respectively. Regression mean square error was 0.02+/-0.04 and 0.09 for training and validation groups respectively (Fig.2A).