Convolutional neural network based brain MRI segmentation

Baydar, Bora
Visualization of the inner parts of human body is crucial in modern medicine and magnetic resonance imaging(MRI) is one of the widely used medical imaging methods. Manual analysis of MRIs, however, wastes the valuable time of experts. Development of an automatic segmentation method for brain MRIs can save time spent by the experts and can avoid human error factor. In this thesis, convolutional neural network (CNN) based methods are applied on brain MRI segmentation problem. The basic architectures used are FCN-8 and U-NET. Performance of different approaches has been analyzed by focusing on structural modifications, upsampling methods, activation functions, loss functions, pre-processing and post-processing methods. For the activation functions, ReLU, LReLU, PReLU and tanh are experimented. Histogram matching, normalization and histogram equalization have been applied for pre-processing. Conditional random fields and a 3 dimensional connected component analysis are separately integrated to the network as post-processors. Results are compared in terms of dice score, sensitivity and specificity. The experimental results show that the combination of two separate U-Nets has the best performance. Dilation modules also improve the results when inserted on a shal-low network. When combined with additional residual connections, they have alsoimproved the overall results. Inception modules do not provide a remarkable performance improvement. ReLU and PReLU have shown the best performance. No significant difference have been observed between results obtained from different upsampling methods, although bilinear interpolation, despite being non-trainable, has slightly better results.