Artificial Neural Networks to Evaluate the Boron Concentration Decreasing Profile in Blood-BPA Samples of BNCT Patients

GARCÍA-REIRIZ, ALEJANDRO G.; MAGALLANES, JORGE F.; ZUPAN, JURE; LIBERMAN, SARA

Capital Federal, Buenos Aires, Argentina

Congreso; 14th International Congress on Neuron Capture Therapy; 2010

Introduction: The irradiation dose in tumor and healthy tissue of a Boron Neutron Cancer Therapy (BNCT) patient is regulated according to its boron concentration in blood. In most treatments this concentration is experimentally determined before and after irradiation, but not during irradiation because it is troublesome to take the blood samples when the patient remains isolated in the irradiation room. A few models are used to predict the boron profile during that period which involves biexponential decay.   Materials and Methods: For the prediction of decay concentration profiles of the p-boronophenylalanine (BPA) in blood during BNCT treatment a method is suggested based on Kohonen neural networks. The results of a (20×20×40) Kohonen network model trained with the data set of 64 concentration profiles extracted from published studies are described and discussed. To make the data comparable among themselves, the concentration profiles were interpolated 40 points in 10 minutes interval between 10 and 400 minutes, taking into account that all profiles have the first point at zero time. The predictions of the concentrations in the range from 30 and 90 minutes after the peak concentration of the fructose-BPA (F-BPA) infusion, up to 400 minutes after the starting points are obtained if the first 20 to 30 % of points of the whole concentration profile are inputted, i.e. the concentrations from time zero to the concentration reached in both sceneries 30 or 90 minutes after the peak occurrence. The prediction results of the method were tested for different types of data set-ups and learning strategies.   Results and Discussions: The prediction ability and robustness of the modeling method were tested by the leave-one-out procedure. It means to run the program every time for a set of 63 profiles, leaving out of the training procedure a different concentration profile and then calculating the prediction error of the omitted profile. The results show that the method is very robust and mostly independent on small variations. In order to show the abilities and limitations of the method the best and the few worst results are discussed in detail. The boron concentration measured after the irradiation of the patient ends, allowed to rebuild the retrospective boron decay profile. The prediction ability to fit the retrospective experimental data curve shows an uncertainty lower than the biexponential approach.  Another advantage of the method that should be emphasized is that by increasing the number of data included in the model it improves automatically the prediction ability and the robustness.