Automatic hippocampal subfield segmentation in ultra-high resolution MRI using Convolutional Neural Network

Jacobsen, Nina ; Nøhr, Anne Krogh ; Hansen, Bolette Dybkjær

4. term

Biomedical Engineering and Informatics, Master

2017

2017-06-06

Objective: The hippocampal subfields are of great interest within research due to a diagnostic connectivity to neurodegenerative diseases, and increased access to ultra-high field MR imaging has made segmentation of these subfields feasible. Automatic methods for segmentation of the hippocampal subfields have been proposed. However, these methods show limitations with respect to segmentation of smaller subfields and they are very time consuming. In current research, Convolutional Neural Networks (CNN) have been used to segment brain structures and lesions, and the approach have shown fast and accurate segmentations. The aim of this study was to develop a CNN based automatic method for hippocampal subfield segmentation (DeepHSS) and explore the impact of a small dataset. Method: DeepHSS consists of a preprocessing pipeline followed by a 12 layer, two pathways 3D CNN trained using dense training. The training set includes 7T TSE MR images acquired from 15 subjects and associated manual labels delineating six hippocampal subfields and Entorhinal Cortex. DeepHSS was tested with 10 subjects. The segmentations by DeepHSS were compared to the corresponding manual labels using the Dice Similarity Coefficient (DSC). Results: DeepHSS showed fast segmentation of hippocampal subfields with an average foreground DSC at 0,83±0.03. Highest average segmentation DSC was achieved for Subiculum (0.80±0.05) and Dendate Gyrus (0.77±0.05) whilst the lowest average segmentation accuracy was found for Entorhinal Cortex (0.49±0.14). The accuracy of the segmentation increased with the number of training subjects. Conclusion: Using our automatic hippocampal subfield segmentation method, DeepHSS, we demonstrated CNNs as an efficient method for automatic hippocampal subfield segmentation despite utilisation of a small dataset. The results were comparable with results obtained using ASHS.

CNN ; convolutional neural network ; segmentation ; MRI ; hippocampal subfields ; ultra-high field ; deep learning

Download
View record in AAU Student Projects