Program Information
A Multilayer Perceptron Model to Predict the Seed Number for Real-Time Low-Dose Rate Prostate Brachytherapy
R Pappafotis1*, D Nie2 , R Chen3 , D Shen4 , L Potter5 , A Price6 , A Wang7 , J Lian8 , (1) Duke University, Durham, North Carolina, (2) The Universitty of North Carolina, Chapel Hill, North Carolina, (3) University of North Carolina, Chapel Hill, NC, (4) The Univerity of North Carolina, Chapel Hill, North Carolina, (5) The University of North Carolina, Chapel Hill, North Carolina, (6) University of North Carolina, Chapel Hill, NC, (7) The Univerity of North Carolina, Chapel Hill, North Carolina, (8) Univ North Carolina, Chapel Hill, NC
Presentations
TU-FG-605-8 (Tuesday, August 1, 2017) 1:45 PM - 3:45 PM Room: 605
Purpose: To use knowledge of patient data to create a model that can accurately predict the quantity of radioactive seeds needed for real-time low-dose rate permanent prostate brachytherapy.
Methods: Fifty-one clinical plans with automatically optimized and manually adjusted seed numbers and positions are retrospectively included in this study. Multilayer perceptron (MLP) has proven to be a reliable machine learning tool to build nonlinear correlation between inputs and outputs. Two 5-layer MLP models were created for predicting the seed number: 1) Model A employed seed activity, prostate volume, prostate coverage, and three prostate dimensions as input parameters; 2) Model B excluded the information of three prostate dimensions. The output of both models was the seed number for the clinical plan. Data from all 51 patients was gathered to train the models using a leave-one-out cross-validation technique (LOOCV). As per LOOCV, each model was trained with 50 patients and tested its accuracy on the remaining patient. The model predictions and the values from a commonly used empirical formula were compared with the true number of seeds. The prediction error band was defined as the seed number within ±10% of the true value.
Results: The empirical formula always overestimated the seed number with an average absolute error of 11.0 and an average percent error of 22.8%. MLP models improved the accuracy of prediction. Specifically, the average absolute error and the average percent error of predictions for the variations of Model A were 1.8 and 4.0%, respectively. The prediction accuracy was 92.2%. The simplified Model B reported a similar performance with slightly increased prediction error.
Conclusion: The MLP can more accurately predict the number of brachytherapy seeds needed for treatment than the current empirical formula. The cost of brachytherapy could be reduced due to fewer excess seeds ordered for the treatment.
Contact Email: