Efficient 3D Neural Networks with Support Vector Machine for Hippocampus Segmentation

摘要

Accurate segmentation of the hippocampal and its subfields from the brain magnetic resonance imaging (MRI), which is a prerequisite for volume measurement, plays a significant role in the clinical diagnosis and treatment of many neurodegenerative diseases. It is of great significance for the precise segmentation of the hippocampus and its sub-regions.In this paper, we proposed a hippocampal subfields segmentation approach based on support vector machine (SVM) combined 3D convolutional neural network (3D CNN) and generative adversarial network (GAN). In the 3D CNN-SVM model, the representative features processed by the 3D CNN are input into the SVM. SVM is trained with the features to achieve the voxel classification of the image, and the segmentation results are obtained. In the 3D GAN-SVM model, we use the generator to segment and use the 3D CNN-SVM network we proposed as the discriminator.The experiments has performed on the dataset obtained from Center for Imaging of Neurodegenerative Diseases (CIND) in San Francisco, USA. The segmentation dice similarity coefficients (DSCs) of the 3D CNN-SVM for CAI, CA2, DG, CA3, Head, Tail, SUB, ERC and PHG in hippocampal subfields are respectively 0.930, 0.926, 0.977, 0.967, 0.931, 0.905, 0.981, 0.870 and 0.911. It demonstrates that combining 3D CNN and SVM achieves a significant improvement in the accuracy of all the hippocampal subfields, and outperforms the existing methods based on the CNN. The DSCs of 3D GAN-SVM are higher, which are respectively 0.989, 0.965, 0.986, 0.977, 0.975, 0.993, 0.818, 0.985 and 0.994. The effect of the GAN-SVM model is also significantly better than that of pure GAN, and the segmentation accuracy has reached the best level on this dataset.Neural network can extract representative features, but it mainly relies on extracting features from a large number of accurately labeled datasets. Most medical datasets are small and difficult to obtain. SVM is more suitable for classification of small datasets, so we combine SVM and neural network to effectively improve the segmentation accuracy of hippocampus in brain MRI images.