Abinash Pant

Electroencephalography (EEG) is a non-invasive and widely used technique to measure electric potential differences that result from current flow in active brain regions. Localization of the relevant brain regions causing the effects on the scalp surface cannot be performed unequivocally due to the so-called “inverse problem”, i.e. different underlying brain activity patterns can cause the same EEG representation. One important factor for identifying the truly active brain regions is having very good knowledge about the so-called “forward model”. As electric current is diffracted differently and reflected to varying degrees by different tissues, it is important to know the anatomy of the head, as well as the conductivity properties of the head tissues as precisely as possible.

Head models representing the conductivity distribution inside the subject's or patient's head play an important role for the solution of the forward, and thus, also the inverse problem. Realistic models taking into account the individual anatomy of the subject can greatly improve the accuracy of source analysis.

Generation of the head model can be accurately estimated with the help of structural imaging techniques. These structural images can be used to segment the different brain regions. An accurate head model could be obtained by using these images to estimate the underlying structure by segmenting the different brain regions. Pipeline to generate head models using segmentation is particularly challenging for children; smaller anatomical structures, motion artifacts, and an anatomy which varies greatly for children at different ages require the development of new algorithms for an accurate and robust segmentation. Development will focus on Bayesian segmentation approaches representing the typical anatomy using Markov Random Field models. Furthermore, age-specific a-priori knowledge on the anatomy of the children's heads will be taken into account.

The major objective of this study is to devise an efficient and accurate method for segmenting the hippocampus. The hippocampus being a small part of the brain is quite hard to segment. An accurate segmentation would however make it possible to measure its volume which, in turn, can be used as a biomarker to detect and track the progression of Alzheimer's disease in a patient's brain.

This work provides an extensive study of existing methods for segmenting the hippocampus. Inferences and observations from this perusal of existing methods were used to design a robust and accurate segmentation method. A multi-atlas based method was developed to segment the hippocampus, one patch at a time. This was achieved using a reconstruction algorithm based on sparse representation. Furthermore, a smaller feature representation is used along with hierarchical pruning to include more priori, resulting in an accurate segmentation. The work also presents ways of optimizing the method to make it computationally effective by introducing clever heuristics and leveraging existing approaches.

This method was validated using open source datasets and provided an average Dice score of (L:0.892,R:0.893) using affine registration, which is on par with the state of the art. A thorough validation also proved the effectiveness of the proposed improvements. A study was also undertaken to determine the effectiveness of the method to produce accurate volume. The volume computed using the proposed method was close to the actual volume with an average deviation of 3.87% .