Electronic Thesis and Dissertation Repository

Thesis Format

Integrated Article


Doctor of Philosophy




Khan, Ali R.

2nd Supervisor

Köhler, Stefan



The hippocampal subfields are defined by their unique cytoarchitectures, which many recent studies have tried to map to human in-vivo MRI because of their promise to further our understanding of hippocampal function, or its dysfunction in disease. However, recent anatomical literature has highlighted broad inter-individual variability in hippocampal morphology and subfield locations, much of which can be attributed to different folding configurations within hippocampal (or archicortical) tissue. Inspired in part by analogous surface-based neocortical analysis methods, the current thesis aimed to develop a standardized coordinate framework, or surface-based method, that respects the topology of all hippocampal folding configurations. I developed such a coordinate framework in Chapter 2, which was initialized by detailed manual segmentations of hippocampal grey matter and high myelin laminae which are visible in 7-Tesla MRI and which separate different hippocampal folds. This framework was leveraged to i) computationally unfold the hippocampus which provided implicit topological inter-individual alignment, ii) delineate subfields with high reliability and validity, and iii) extract novel structural features of hippocampal grey matter. In Chapter 3, I applied this coordinate framework to the open source BigBrain 3D histology dataset. With this framework, I computationally extracted morphological and laminar features and showed that they are sufficient to derive hippocampal subfields in a data-driven manner. This underscores the sensitivity of these computational measures and the validity of the applied subfield definitions. Finally, the unfolding coordinate framework developed in Chapter 2 and extended in Chapter 3 requires manual detection of different tissue classes that separate folds in hippocampal grey matter. This is costly in the time and the expertise required. Thus, in Chapter 4, I applied state-of-the-art deep learning methods in the open source Human Connectome Project MRI dataset to automate this process. This allowed for scalable application of the methods described in Chapters 2, 3, and 4 to similar new datasets, with support for extensions to suit data of different modalities or resolutions. Overall, the projects presented here provide multifaceted evidence for the strengths of a surface-based approach to hippocampal analysis as developed in this thesis, and these methods are readily deployable in new neuroimaging work.

Summary for Lay Audience

One way that scientists have studied the human hippocampus is through Magnetic Resonance Imaging (MRI). While non-invasive, MRI has limitations in resolution and contrast and so the current thesis aims to gain as much information as possible from such images. Given only a blurry image of the hippocampus, we can get a rough idea of its overall shape and properties, which are often estimated in current research and clinical examinations. From existing literature, we know that the hippocampus is composed of a thin, folded sheet of tissue. Thus, there may only be a few possible folding patterns that could produce a given coarse hippocampal shape and image. The current thesis tries to reconstruct possible folding patterns of the hippocampus from MRI or other images.

Once we understand how a given hippocampus is folded, it becomes easier to learn more about its structure. For example, we can measure its thickness and other properties, or we can more easily divide it into contiguous subfields. For easier visualization, we can also computationally ‘unfold’ this structure and plot its properties, such as thickness, across its full 3D extent in just one flattened plane of view. Doing this type of unfolding also allows us to align different hippocampi despite differences in their original, 3D folded shapes. This can be used to align many hippocampi and look for subtle statistical differences between subgroups.

A computational approach to understand hippocampal folding is presented in Chapter 2 of this thesis, but this still relies on some manual input. In Chapter 3 we extend these methods to a unique 3D histological dataset and show that our computational unfolding approach alone can be used to detect hippocampal subfields, instead of more traditional detection by trained neuroanatomists. In Chapter 4 we fully automate the application of this computational unfolding to new MRI data using deep learning instead of manual delineation of tissues separating different folds of the hippocampus. Altogether, these methods could help us identify which properties of the hippocampus are correlated with performance on tasks, like episodic memory tests, and which properties are correlated with, or diagnostic of, neurological diseases.

Creative Commons License

Creative Commons Attribution 4.0 License
This work is licensed under a Creative Commons Attribution 4.0 License.