Evolutionary learning of spiking neural networks towards quantification of 3D MRI brain tumor tissues

This paper presents a new classification technique for 3D MR images, based on a third-generation network of spiking neurons. Implementation of multi-dimensional co-occurrence matrices for the identification of pathological tumor tissue and normal brain tissue features are assessed. The results show the ability of spiking classifier with iterative training using genetic algorithm to automatically and simultaneously recover tissue-specific structural patterns and achieve segmentation of tumor part. The spiking network classifier has been validated and tested for various real-time and Harvard benchmark datasets, where appreciable performance in terms of mean square error, accuracy and computational time is obtained. The spiking network employed Izhikevich neurons as nodes in a multi-layered structure. The classifier has been compared with computational power of multi-layer neural networks with sigmoidal neurons. The results on misclassified tumors are analyzed and suggestions for future work are discussed.