Image analysis by pulse coupled neural networks (PCNN)—a novel approach in granule size characterization

Abstract

A biologically inspired spiking neural network model, the pulse coupled neural network (PCNN), has been applied for the first time in bulk particle characterization, and specifically in the characterization of pharmaceutical granule size distributions. The PCNN was trained on surface images of pharmaceutical granule beds, and the adjustable parameters (radius neuron interconnection, r0, linking weight coefficient, β, local threshold potential, VΘ, and number of iterations) were successfully optimized using design of experiments. As demonstrated with size fractions of granules, it was found that the PCNN produced granule size-dependent signals. In general, a first highest and relatively narrow peak located in the region of two to twelve iterations corresponded to smaller particle size, while larger particles resulted in wider peaks and in highest (not first) peak at a range between 13 and 25 iterations. Better predictions, i.e. lower RMSEP (root mean squared error of prediction) values, were obtained using high β value, low r0 and VΘ values, while the number of iterations had to exceed 110 and the optimized model (RMSEP lower than 5) corresponded to PCNN variables: r0 = 1, β = 0.4, VΘ = 2, and number of iterations = 150. The coefficient of determination (R2) of the model was 0.94 and the predicted variation (Q2) was 0.91, while the Pearson correlation coefficient between the predicted and the measured mean particle size by sieving for eight test batches was 0.98. These findings could be characterized as promising and encouraging for the further use of image analysis by PCNNs in pharmaceutical bulk particle size and shape characterization.