Design of a Convolutional Neural Network for Hippocampal Segmentation in Epileptics and Healthy Patients

Manual segmentation is still the gold standard in hippocampal segmentation due to its precision. However, it requires a considerable amount of time. Convolutional neural networks offer a less resource-intensive alternative. In this study, we propose a parallel convolutional neural network architecture for segmenting the hippocampus in patients with epilepsy and healthy patients based on magnetic resonance images. Our network design resembles a wavelet filter bank but utilizes adaptive filters generated by convolutional layers. By employing a limited number of convolutional layers, our approach achieves improved computational efficiency compared to existing network models in the literature. The performance evaluation was conducted using the Jaccard index, Dice coefficient, sensitivity, and precision, and compared to the widely used U-Net network segmentation. The results, based on similarity metrics, indicate that the parallel network demonstrates higher predictive similarity to the test dataset, achieving a precision score of 0.80 outperforming the U-Net network.