A multilayer perceptron (MLP) neural network is built to analyze the Cs-137 concentration in seawater via gamma-ray spectrums measured by a LaBr3 detector. The MLP is trained and tested by a large data set generated by combining measured and Monte Carlo simulated spectrums under the assumption that all the measured spectrums have 0 Cs-137 concentration. And the performance of MLP is evaluated and compared with the traditional net-peak area method. The results show an improvement of 7% in accuracy and 0.036 in the ROC-curve area compared to those of the net peak area method. And the influence of the assumption of Cs-137 concentration in the training data set on the classifying performance of MLP is evaluated.