About the project
Epilepsy is a prevalent disorder but much about it is still not understood. Many cases originate from regions in the hippocampus and surrounding regions. However, its complex structure makes it hard to research.
My project was to use a new highly detailed protocol for defining subfields of the hippocampus to automatically segment it. It was hypothesised that a convolutional neural network (CNN) would yield results that rivaled the accuracy of literature methods.
I implemented the SegNet architecture and trained the network on a dataset of 7T MRI images. The trained network was able to generate 2D predictions of subfields in images. I stacked the 2D images into a 3D array reconstruct each of the coronal, axial, and saggital views for clinical use.
All data analysis and visualization was done in Matlab. The results were highly variable, performing better than literature methods in images where features where large, but very poorly where features in images were small. It is believed that a larger dataset along with a pre-trained network would further improve these results.
Predictions generated by trained model of grey matter label and alternate orientations after reconstruction.