Elsevier

NeuroImage

Volume 163, December 2017, Pages 286-295
NeuroImage

HIPS: A new hippocampus subfield segmentation method

https://doi.org/10.1016/j.neuroimage.2017.09.049Get rights and content

Highlights

  • We present a novel method for hippocampus subfield segmentation on MRI.

  • The method consists of a fast multi-atlas non-local patch-based label fusion.

  • Our proposed method was shown to improve the state-of-the-art methods with a reduced temporal cost (20 mins).

  • The pipeline presented in this work will be made available to scientific community through our web -based platform volBrain.

Abstract

The importance of the hippocampus in the study of several neurodegenerative diseases such as Alzheimer's disease makes it a structure of great interest in neuroimaging. However, few segmentation methods have been proposed to measure its subfields due to its complex structure and the lack of high resolution magnetic resonance (MR) data. In this work, we present a new pipeline for automatic hippocampus subfield segmentation using two available hippocampus subfield delineation protocols that can work with both high and standard resolution data. The proposed method is based on multi-atlas label fusion technology that benefits from a novel multi-contrast patch match search process (using high resolution T1-weighted and T2-weighted images). The proposed method also includes as post-processing a new neural network-based error correction step to minimize systematic segmentation errors. The method has been evaluated on both high and standard resolution images and compared to other state-of-the-art methods showing better results in terms of accuracy and execution time.

Introduction

The hippocampus (HC) is a small bilateral brain structure located in the medial temporal lobe at both sides of the brainstem near to the cerebellum. Its name comes from its similarity to the sea-horse. Starting from the upper end at the hippocampal sulcus we find the dentate gyrus (DG) followed by the Cornu Ammonis (CA) which is subdivided in four consecutive parts (CA4 to CA1) and the Subiculum at the bottom end. The CA is also structured in six layers called strata. These layers are the Stratum oriens (SO), Stratum pyramidale (SP), Stratum lucidum (SLU), Stratum radiatum (SR), Stratum lacunosum (SL) and the Stratum molecuare (SM).

HC is involved in many brain functions such as memory and spatial reasoning (Milner, 1958; Schmajuk, 1990, Hafting et al., 2005). Several studies showed that it has an important role in many neurodegenerative diseases such as Alzheimer's disease (AD) (Braak and Braak, 1991) or schizophrenia (Altshuler et al., 1998). The study of the hippocampus volume is of great interest as it is a valuable tool for follow-up and treatment adjustment (Jack et al., 2000, Jack et al., 2005, Dickerson and Sperling, 2005). However, the HC anatomy is complex and variable, and the limits between different subfields have been described in the neuroanatomy literature using cytoarchitectonic features that require histological staining and microscopic resolution to visualize (Insausti and Amaral, 2004).

Due to the key importance of this structure, several segmentation methods and protocols have been developed (Barnes et al., 2008, Collins and Pruessner, 2010, Coupé et al., 2011). However, one of the main problems to advance in this field was the disparity of HC definitions and the lack of manually labelled cases. Recently, a harmonized full hippocampus protocol has been proposed (jointly with 120 1 mm3 resolution manually segmented examples) which will become the common reference for the development and comparison of new segmentation methods (Boccardi et al., 2015). Classically, due to the limitations in MR image resolution, the studies were restricted to consider the hippocampus as a single structure (Chupin et al., 2009). Even though the analysis of the whole hippocampus has been shown to be a good approach to study AD, some ex-vivo studies revealed that normal aging and AD affects the subfields differently during the lifespan (Braak and Braak, 1991).

Currently, many HC subfield segmentation protocols have been developed as a response to the advances in MR sequences that allow acquiring high resolution images making possible to divide the hippocampus into its constituent parts. However, there is still little consensus between the different HC subfield protocols as shown in (Yushkevich et al., 2015a) where 21 delineation protocols were compared. Some of these protocols have been used to create anatomically labeled MRI datasets which are a fundamental resource to develop new segmentation methods. For example, 9.4 T ultra-high resolution ex-vivo images were used to create an anatomical atlas (Yushkevich et al., 2009) including the CA1, CA2-3, the DG and the vestigial hippocampal sulcus obtained by manual delineation. In 2013, Winterburn presented a new in-vivo high resolution atlas (Winterburn et al., 2013) to divide the hippocampus in five different subregions: CA1, CA2-3, CA4/DG, Stratum and Subiculum (jointly with 5 manually segmented examples, we call this the Winterburn dataset). Later in 2015, Kulaga-Yoskovitz developed another segmentation protocol (Kulaga-Yoskovitz et al., 2015) consisting of three structures: CA1-3, CA4/DG and Subiculum (jointly with 25 manually segmented examples, we call this the Kulaga-Yoskovitz dataset).

To conduct volumetric studies and apply these delineation protocols, automatic segmentation tools are necessary. It is well known that manual delineation of a new case represents an issue in terms of reproducibility. It is also extremely time consuming as well as it has a high economic cost (it can take from 10 to 20 h of an expert rater time per subject to manually segment the hippocampus subfields (Iglesias et al., 2015)). Since manual segmentation is not an affordable option, several automatic methods have been developed in the last years. One of the first HC subfield segmentation methods was proposed by Van Leemput (Van Leemput et al., 2009) using a generative model of the hippocampus region. This model is produced using a mesh-based probabilistic atlas containing information about where the anatomical labels are most likely to occur. The probabilistic atlas is learned from a set of ultra high resolution training images. Recently, Iglesias (Iglesias et al., 2015) continued this work and improved the model by using a more accurate atlas generated from ultra-high resolution ex-vivo MR images and also using multi-contrast data. Pipitone proposed a multi-atlas-based method (Pipitone et al., 2014) using T2w images intended to segment a considerable large dataset (targets) using a few manually labeled cases (atlases). However, this method, named MAGeT (Chakravarty et al., 2013) has a high temporal cost. In 2015, Yushkevich proposed another method (Yushkevich et al., 2015b) using T2w images where a multi-atlas approach is combined with a similarity-weighted voting and a boosting based error correction. Unfortunately, this method takes hours to produce a segmentation due to the exhaustive use of non-linear registrations as in the case of MAGeT. Recently, in 2016, Caldairou presented a new hybrid method (Caldairou et al., 2016) where a set of training subjects are non-linearly registered to the test case. Then, using patch-correspondences, a surface mesh is generated from the manual labels. These patch correspondences are re-computed for each mesh vertex minimizing the error to adjust a deformable model to the case to be segmented.

In this paper, we propose a new patch-based segmentation method which has been validated using two hippocampus subfield segmentation protocols with publically available datasets. Our method uses an adaptation of MOPAL (Romero et al., 2016), a multi-contrast version of a patch matching segmentation method OPAL (Giraud et al., 2016) to produce fast and accurate segmentations. The presented method works using high resolution (0.5 × 0.5 × 0.5 mm3) T1w and T2w images. It also works on standard resolution images as well as single T1w or single T2w images. During our validation, we show that the proposed approach performs well also on mono-contrast T1w and T2w images as well as when using standard resolution images upsampled using the LASR (Manjón et al., 2010a, Coupé et al., 2013) superresolution method. Our method also includes a new error corrector post processing step based on the use of a boosted ensemble of neural networks is proposed to minimize systematic segmentation errors at post-processing.

Section snippets

Image data

In this work, we have used two different datasets corresponding to two manual labeling hippocampus subfield segmentation protocols, both with high resolution (HR) T1w and T2w MR images. An example of these images and their manual labels can be seen in Fig. 1.

Experiments and results

In this section, the parameters of the proposed method and its results are presented. The method parameters has been adjusted independently to work with the two different segmentation protocols/datasets and it has been compared to other related state-of-the-art methods. To evaluate the segmentation accuracy, we have used the DICE coefficient (Zijdenbos et al., 1994) measured in the linear MNI space. In order to evaluate the significance of the results we applied a Kruskal-Wallis test to find

Discussion

One of the contributions of this work is a new multi-contrast patch similarity consisting on a multi-contrast SSD-based semi-norm (MSN). Introducing this similarity measure in OPAL (now MOPAL), we achieved good segmentation results using T1w + T2w images. By using the semi-norm to combine distances we obtain a robust and self-balanced similarity that takes benefit from information coming from both channels. This means image corruption or low image quality in one of the channels can be overcome

Conclusion

In this work, we have presented a new method for HR hippocampus subfield segmentation called HIPS. It uses two publically available segmentation protocols and datasets (Winterburn and Kulaga-Yoskovitz). Our method is based on MOPAL, a multi-contrast extension of the OPAL patch-based label fusion segmentation method and a novel neural network based error corrector. HIPS works in a fully automated manner providing accurate results in less than 20 min thanks to MOPAL that performs fast

Acknowledgements

This research was supported by Spanish UPV2016-0099 and TIN2013-43457-R grants from UPV and the Ministerio de Economia y Competitividad. This study has been carried out with financial support from the French State, managed by the French National Research Agency (ANR) in the frame of the Investments for the future Program IdEx Bordeaux (ANR-10-IDEX-03-02, HL-MRI Project), Cluster of excellence CPU and TRAIL (HR-DTI ANR-10-LABX-57) and the CNRS multidisciplinary project ”Défi imag'In“. We also

References (39)

Cited by (55)

  • Memory decay distinguishes subtypes of gist

    2021, Neurobiology of Learning and Memory
  • Hippocampal subfield volumes across the healthy lifespan and the effects of MR sequence on estimates

    2021, NeuroImage
    Citation Excerpt :

    Here, we randomly chose four participants (two females and two males) per 5 year bin. HIPpocampus subfield Segmentation (HIPS; (Manjon et al., 2020; Romero et al., 2017) was applied to all the T1w scans of our datasets to investigate the differences of segmentation technique with MAGeT. We decided to investigate this other segmentation technique since it uses the same atlases as MAGeT (Winterburn et al., 2013).

View all citing articles on Scopus
View full text