Koen Van Leemput1,2, Akram Bakkour1,3, Thomas Benner1, Graham Wiggins1,

1
Automated Segmentation of Hippocampal Subfields
from Ultra-High Resolution In Vivo MRI
Koen Van Leemput1,2, Akram Bakkour1,3, Thomas
Benner1, Graham Wiggins1, Lawrence L. Wald1,4,
Jean Augustinack1, Bradford C. Dickerson1,3,
Polina Golland2, and Bruce Fischl1,2,4
1 Athinoula A. Martinos Center for Biomedical
Imaging, Department of Radiology, MGH, Harvard
Medical School, USA 2 Computer Science and
Artificial Intelligence Laboratory, MIT, USA 3
Department of Neurology, MGH, Harvard Medical
School, USA 4 Harvard-MIT Division of Health
Sciences and Technology, MIT, USA
2
Introduction

• The hippocampus consists of multiple, interacting
  subregions
• Distinct hippocampal subregions
• are implicated in different memory subsystems
• are differentially affected in different
  conditions and disease processes (normal aging,
  Alzheimers disease (AD), )

3
Introduction
The ability to reliably detect these subregions
using in vivo neuroimaging is of great potential
value Basic neuroscience insights into the
function and structure of the hippocampus in the
living human brain, and how it is affected in
normal aging. Clinical research sensitive,
non-invasive biomarkers for early diagnosis,
treatment evaluation, and clinical trials in AD.
4
Introduction

• Recent developments in MR data acquisition
  technology are starting to yield images that show
  anatomical features of the hippocampal formation
  at an unprecedented level of detail

New opportunities for explicitly quantifying
individual subregions, rather than their
agglomerate, directly from in vivo MRI
Standard resolution 1 x 1 x 1 mm3
Analyzing large imaging studies of ultra-high
resolution MRI scans requires automated
computational techniques
Ultra-high resolution 0.38 x 0.38 x 0.8 mm3
5
Model-based segmentation

• Derive computational models from manual
  delinations in a number of subjects
• Use those models to automatically segment MRI
  scans of new subjects

Manual delineation in subject 1
Manual delineation in subject N

Scan of new subject
6
Probabilistic atlas

Manual delineation in subject 1
Manual delineation in subject N

prior probability for subiculum
prior probability for white matter
etc
7
Probabilistic atlas
8
Statistical model of image formation
prior probability for white matter
prior probability for subiculum
etc
9
Statistical model of image formation
deformed prior probability for white matter
deformed prior probability for subiculum
etc
10
Statistical model of image formation
Label image
etc
11
Statistical model of image formation
Label image
MR image
etc
12
Segmentation solving the inverse problem
MR image
13
Segmentation solving the inverse problem
MR image
Label image
14
Segmentation solving the inverse problem
MR image
Label image
15
Segmentation solving the inverse problem
MR image
Label image
Bayesian inference

• Start from our statistical model of image
  formation
• Play with the mathematical rules of probability

16
Example
17
Example
18
Example
19
Example
20
Example
21
Example
22
Example
23
Validation Van Leemput et al., Hippocampus
2009

• Ultra-high resolution (0.38 x 0.38 x 0.8 mm3) MRI
  data
• 3T Siemens Trio with prototype custom-built
  32-channel coil
• Optimized MPRAGE sequence, 208 coronal slices
• 5 acquisitions motion-corrected and resampled to
  0.38mm isostropic
• Manual segmentation of the right hippocampus in
  10 subjects (6 younger 4 older cognitive
  normal, age range 22-89)
• Fimbria, CA1, CA2/3, CA4/DG, presubiculum,
  subiculum, hippocampal fissure surrounding
  structures
• Extremely time consuming several days per
  hippocampus!

Validation of the automated segmentation
algorithm using leave-one-out cross-validation
24
Qualitative results
MRI data
manual segmentation
automated segmentation
25
Quantitative results spatial overlap

• Dice coefficient for automated vs. manual
  segmentation
• ( volume of overlap ) / ( average volume )

26
Quantitative results spatial overlap

• Two slices in the midbody of the hippocampus
  re-segmented twice by the same operator for five
  subjects
• Dice overlap between repeated manual
  segmentations (empty bars) vs. Dive overlap
  between automated and manual segmentations
  (filled bars)

27
Quantitative results volume measurements

• Volume differences between manual and automated
  segmentations
• Linear regression between volume measurements
  with both methods (Pearsons correlation
  coefficient)

0.91 p 0.0002
0.60 p 0.066
0.83 p 0.0028
Corr. coeff.
28
Currently working on

• Segmenting hippocampal subfields in lower
  resolution scans
• Explicitly modeling the partial volume
  effect

Standard resolution 1 x 1 x 1 mm3
29
Currently working on

• Segmenting hippocampal subfields in standard
  resolution (1x1x1mm3) scans
• Quantifying segmentation
  uncertainties
• (error bars on
  measurements)

30
Currently working on

• Incorporating the software code into FreeSurfer