Semantic Segmentation of Hippocampal Subregions With U-Net Architecture

Semantic Segmentation of Hippocampal Subregions With U-Net Architecture

Soraya Nasser (Université Oran 1, Algeria), Moulkheir Naoui (Université Oran 1, Algeria), Ghalem Belalem (Université Oran 1, Algeria) and Saïd Mahmoudi (Mons University, Belgium)
Copyright: © 2021 |Pages: 20
DOI: 10.4018/IJEHMC.20211101.oa4
Article PDF Download
Open access articles are freely available for download

Abstract

The Automatic semantic segmentation of the hippocampus is an important area of research in which several convolutional neural networks (CNN) models have been used to detect the hippocampus from whole cerebral MRI. In this paper we present two convolutional neural networks the first network ( Hippocampus Segmentation Single Entity HSSE) segmented the hippocampus as a single entity and the second used to detect the hippocampal sub-regions ( Hippocampus Segmentation Multi Class HSMC), these two networks inspire their architecture of the U-net model. Two cohorts were used as training data from (NITRC) (NeuroImaging Tools & Resources Collaboratory (NITRC)) annotated by ITK-SNAP software. We analyze this networks alongside other recent methods that do hippocampal segmentation, the results obtained are encouraging and reach dice scores greater than 0.84
Article Preview
Top

Introduction

The hippocampus is a structure that is anatomically complex. This consists of two cortex coiled around each other with gray matter, cornu ammonis (CA) and dentate gyrus (DG), separated by the hippocampal groove (Henry, et al., 2011). The Cornu ammonis is also composed of four CA areas (1-4) each one is composed of different variable layers in cell composition: the Pyramidal Stratum (SP) is richer in the neuronal bodies than the Radiatum (SR) Strata, Lacunosum and Moleculare (SRLM) are poorer in the neuronal bodies (Cury., 2015); (Boutet, et al., 2014). he subiculum is the hippocampal formation's most subordinate component. This sits between the normal hippocampus entorhinal cortex and the normal hippocampus subfield CA1 (see Figure 1).

Figure 1.

Hippocampus anatomy (Györfi et al., 2017)

IJEHMC.20211101.oa4.f01

Hippocampus function is important in cognitive processes such as short-term processing of memory, thinking, spatial navigation and behavioral inhibition. The hippocampus is one of the first structures affected in Alzheimer's disease (Bobinski et al., 1999), epilepsy and schizophrenia (Koolschijn et al., 2010). It is characterized by atrophies in its shape, the volume of which would be a risk factor. This atrophy of the hippocampus can be assessed from magnetic resonance imaging (MRI) using manual segmentation like in (Kerchner, et al., 2010); (Henry, et al., 2011); (Boutet, et al., 2014). The hippocampus was manually segmented according to the anatomy described by (Duvernoy, 2005). The sub-regions are the Ammon horn, specifically the CA1, CA2, CA3 fields, CA4 and the dentate gyrus (DG) (Figure 1).

Wisse et al. (2012, 2014) proposed a new manual segmentation protocol for hippocampal sub-regions.

However, a lot of work has been done to develop manual segmentation to an automatic segmentation that provides a quantitative and qualitative estimate of this complex structure who they include the notion of registration and deformation using atlas(es) (Chupin, et al., 2009; Kim et al., 2011).

In contrast, an effort was made to systematically analyze the anatomical variation of the hippocampus from large data sets by estimating the iterated barycenter of the population as a model for the analysis of anatomical variability by principal component analysis (PCA) on initial moment vectors or an approximate distance matrix (Cury, 2015).

Our research aims at achieving an automatic semantic segmentation evaluating the hippocampus as a single entity from the MRI on the one side and the hippocampal sub-regions on the other.

This paper will be organized as follows, we review two states-of-the-art, the first concerning the semi-automated and automated hippocampus segmentation and the second one in the field of medical images segmentation using CNNs model (Related work section). Then we detail our approach via two convolutional neural networks (Method section). Finally, the results are reported and discussed (Experiment, Results sections). The conclusion and perspective are described at the end (Conclusion section).

Top

The diagnosis is performed in a clinical context; the study of the morphological features of the hippocampus is based on a visual examination by an often biased neuro-radiologist.

On the other hand, semi-automated segmentation approaches are more advanced, such as those based on deformable models and the use of atlases. Both approaches use precompiled geometric information that can be created from datasets on real subjects. User interaction with these strategies only occurs at the initial stage, which significantly eliminates manual activity.

(Ashton et al., 1997) proposed a method based on Active Contour Model (ACM).

Complete Article List

Search this Journal:
Reset
Volume 13: 2 Issues (2022): 1 Released, 1 Forthcoming
Volume 12: 6 Issues (2021)
Volume 11: 4 Issues (2020)
Volume 10: 4 Issues (2019)
Volume 9: 4 Issues (2018)
Volume 8: 4 Issues (2017)
Volume 7: 4 Issues (2016)
Volume 6: 4 Issues (2015)
Volume 5: 4 Issues (2014)
Volume 4: 4 Issues (2013)
Volume 3: 4 Issues (2012)
Volume 2: 4 Issues (2011)
Volume 1: 4 Issues (2010)
View Complete Journal Contents Listing