Hippocampus Segmentation Using Locally Integrated Prior-Based Level Set Guided By Assembled And Weighted Coherent Point Drift Registration
Loading...
Date
2016-03
Authors
Achuthan, Anusha
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
Hippocampus segmentation from neighbouring brain subcortical structures is
a very challenging task mainly because boundaries separating these structures are
weak or unclear, rendering conventional edge-based approaches ineffective for proper
hippocampus segmentation. Besides that, close proximity of the hippocampus
with the amygdala further complicates the segmentation issue. Recent trends,
however have shifted from sole reliance on image features to utilization of prior
models in the segmentation. Predominantly, the prior models are constructed using
atlas-based segmentation. This approach however, is highly data intensive due to the
volumetric-based methods used for prior model construction. Consequently, this thesis
proposes a prior model construction method that not only effectively represents shape
and spatial location information, but also requires lower data intensiveness compared to
atlas-based approaches. Specifically, a novel point set registration method is proposed
and validated for prior model construction. Instead of using the whole image volume,
the proposed point set registration utilizes a set of representative points in constructing
the prior model. This leads to the next contribution of this thesis where a locally
integrated prior-based level set is introduced for final hippocampus segmentation. The
locally integrated prior-based level set used the prior model only at locations with insufficient boundary information for accurate segmentation. This is the main key
feature compared to previously proposed approaches that perform global integration
of the prior information, that employed prior model throughout the image domain.
Evaluations on the constructed prior model were carried out using the OASIS-MICCAI
dataset. Compared to the more popular Mean Shape in Signed Distance Map approach,
the proposed prior model construction approach showed improvement by 1.59% in
average Root Mean Square Error, especially in generalizing target hippocampus that
does not fall within a training population. It is also demonstrated that the prior model
construction is less data intensive compared to atlas-based approaches, in terms of
number of data points being used during the construction. Final segmentation results
indicate that the proposed locally integrated prior-based level set performs better
than the globally integrated prior-based level set, with a 3.36% improvement in Dice
similarity coefficient value. Further comparisons on Dice similarity coefficient have
also shown that the final segmentation results are at par with current state-of-the-art
techniques, outperforming a well known hippocampus segmentation software known
as Freesurfer. Promising improvement shown by the proposed work in this thesis
provide an insight on the applicability of this approach for hippocampus segmentation.
Description
Keywords
Hippocampus