Automatic adjustment of the pulse-coupled neural network hyperparameters based on differential evolution and cluster validity index for image segmentation

Abstract

The pulse-coupled neural network (PCNN) is a cortical model that can be used in image segmentation applications. The performance of the PCNN depends on adjusting its hyperparameters, where population-based metaheuristics, such as evolutionary algorithms, have been used to perform this task by optimizing a fitness function. In this regard, the entropy criterion is a common fitness function used to evaluate the quality of potential PCNN solutions. However, maximizing the entropy is related to maximize the inter-group separation, but the intra-group cohesion is unconsidered. In this regard, a cluster validity index (CVI) can be used as a fitness function, which defines a relationship between inter-group separation and intra-group cohesion. Therefore, we propose using a CVI to quantify the segmentation quality generated by the PCNN given a set of hyperparameters adjusted by the differential evolution algorithm. The proposed approach is tested on a dataset of natural images, where every image has three reference segmentations; thus, the Jaccard index is used to measure the segmentation performance. The experimental results reveal that a simplified PCNN, when used jointly with the Silhouette index, obtains the best performance with a mean Jaccard value of 0.77, whereas the entropy criterion attains 0.41. Additionally, the proposed approach is tested on two modalities of medical images to show its applicability in other kinds of images. The results suggest that using a CVI instead of the entropy criterion can improve the segmentation performance of the PCNN.