5/3/16

1

AUTOMATED HIPPOCAMPUS SEGMENTATION OF BRAIN MRI IMAGES Group 7 Yingchuan Hu Chunyan Wu Li Zhong

May 1st, 2014 Cornell University

Problem Statement • Goal: automated segmentation of hippocampus in brain MRI images

Input Output hippocampus

Hippocampus Anatomy

*Modified from a scan of a plate of “Posterior and inferior cornua of left lateral ventricle exposed from the side” in Gary’s Anatomy

The hippocampus is located in the medial temporal lobe of the brain.

Frontal lobe

Occipital lobe

Temporal lobe

Clinical Significance •  Epilepsy is the most common serious brain disorder worldwide.

•  Prevalence of epilepsy worldwide (WHO1)

•  7 sufferers in every 1,000 people •  3 new sufferers in every 10,000 people each year

•  People with epilepsy are at increased risks for status epilepticus (life-threatening) •  One continuous, unremitting seizure lasting longer than five

minutes or recurrent seizures without regaining consciousness between seizures for greater than five minutes.

•  Prevalence of status epilepticus in US (NIH2) •  195,000 new patients of status epilepticus each year •  42,000 deaths caused by status epilepticus each year

1.  World Health Organization http://www.who.int/mental_health/neurology/epilepsy/en/ 2.  National institute of Health http://www.ninds.nih.gov/disorders/epilepsy/detail_epilepsy.htm

Clinical Research

• Hippocampal volume reduction >10% of “normal” size indicates epilepsy.[1-4]

•  “normal”: •  People with the same age •  Bilateral hippocampus comparison •  Personal changes in more than 1 year

• Over 90% sensitivity + 98% specificity for MRI image measurement diagnosis.[5-7]

1. Cook MJ, et al. (1992). Brain 115:1001–1015. 2. Jack, C. R., et al. (1990). Radiology 175:323–429. 3. Jackson, G. D., et al. (1994). Neurology 44:42–46. 4. Bronen RA, et al. (1995). AJNRAmJ Neuroradiol16:1193–1200. 5. Oppenheim C, et al. (1998). AJNRAmJNeuroradiol 19:457–463. 6. Jack CR. (1995). Neuroimaging Clinics of North America 5:597–622. 7. Tien RD, et al.(1993). Radiology 189:835–842 8. Baulac M., et al. (1998). Ann Neurol 44:223-33

Issues for Segmentation

•  Low contrast to neighboring brain structures

• No clear boundary between hippocampus and amygdala

amgydala hippocampus

5/3/16

2

Previous work

1. Tu, Z. et al. (2008). Brain anatomical structure segmentation by hybrid discriminative/generative models. IEEE Transactions on Medical Imaging, 27, 495-508. 2. Aljabar, P. et al. (2007). Classifier selection strategies for label fusion using large atlas databases. Medical Imaging Computing and Computer-Assisted Intervention, 10, 523-531. 3. van der Lijn, F. et al. (2008). Hippocampus segmentation in MR images using atlas registration, voxel classification, and graph cuts. Neuroimage, 43(4), 708-720. 4. Fiorina, E. et al. (2012, June). Fully automated hippocampus segmentation with virtual ant colonies. In Computer-Based Medical Systems (CBMS), 2012 25th International Symposium on (pp. 1-6). IEEE. 5. Confidence Interval Calculator

Dataset Method Result5

Tu1 2008 Same as our dataset

Hybrid generative/discriminative model,

PCA

[0.61-0.67] Dice Coeff

Aljabar2 2007

Same as our dataset

Multi-atlas model, label fusion classifier

[0.71-0.79] Dice Coeff

Van3 2008 518 cases

(20 manually marked)

Atlas registration, voxel classification

and graph cuts

[0.763-0.902] Dice Coeff

Fiorina4 2012 56 cases Adaptive threshold,

probability map [0.73-0.75] Dice Coeff

99% confidence interval Algorithm Overview

1. Somasundaram et al. Segmentation of hippocampus from human brain MRI using mathematical morphology and fuzzy logic. In Emerging Trends in Science, Engineering and Technology (INCOSET), 2012 International Conference on (pp. 269-273). IEEE.

Somasundaram’s algorithm1:

Issues: no clear boundary & low contrast

Noise removal

Morphologcial

operations Thresholdi

ng Middle-block

selection

Fuzzy logic edge detection

Largest Connected Component

•  Min/max Filtering: a combination of min and max filtering to enhance image contrast

•  min(): min filter max(): max filter f: original image

Our algorithm based on Somasundaram:

Noise removal

Min/max filtering

Thresholding

Middle-block

selection Edge

detection Largest

connected component

filter size w1 filter size w2 block control points P1,P2

Extract peaks

Original image f à f1 = max(min(f)) à peaks fp = f - f1

Extract valleys

Original image f à f2 = min(max(f)) à valleys fv = f2 - f

Enhanced image fe = f + fp - fv

enhanced

original

before min/max filtering after min/max filtering

Noise removal

Min/max filtering

Thresholding

Middle-block

selection Edge

detection Largest

connected component

filter size w1 filter size w2 block control points P1,P2

Properties of min/max filtering:

• More enhancement on “small” peaks •  Separate the hippocampus region from neighboring

regions

• Control enhanced peaks by filter size •  Use proper size of filter to choose peaks for enhancement

• Hippocampus is located in the medial temporal lobe à only consider the middle part of the sagittal view à defined as “middle-block”

• Middle-block •  Rectangle defined by two control

points P1(x1,y1), P2(x2,y2) •  x1=image width / 3 •  x2=2x1

•  y1=image height / 2 – 20 •  y2=image height / 2 + 20

P1

P2

Noise removal

Min/max filtering

Thresholding

Middle-block

selection Edge

detection Largest

connected component

filter size w1 filter size w2 block control points P1,P2

Hypothesis • The algorithm can segment the hippocampus-amygdala region in all cases to 0.7 DC

• The algorithm can segment hippocampus and amygdala for 3.0T cases to 0.7 DC

• Evaluation function: •  Dice Coefficient (DC): 2(A ∩ B) / (A + B) •  A: Ground truth B: Segmentation result

5/3/16

3

Dataset • Data Set1 (with manual markings)

• C1: •  GE 1.5T Coronal T1W MR •  voxel size=0.78x0.78x2mm image dim=256x256x124

• C2: •  GE 3.0T Coronal T1W MR •  voxel size=0.39x0.39x2mm image dim=512x512x112

• 15 C1

•  10 epilepsy(E) & 5 non-epilepsy(N)

• 10 C2 (E) 1.Jafari-Khouzani, K. et al. (2011). Dataset of magnetic resonance images of nonepileptic subjects and temporal lobe epilepsy patients for validation of hippocampal segmentation techniques. Neuroinformatics, 9(4), 335-346.

Experiment •  For C1

•  Training set: 8 C1 (5 E & 3 N) •  Testing set: 7 C1 (5 E & 2 N)

•  For C2 •  Training set: 5 C2 •  Testing set: 5 C2

•  Parameters tuning

Noise removal

Min/max filtering

Thresholding

Middle-block

selection

Largest connected component

Mean filter size w1

Filter size w2

Block control points P1, P2

Changes •  Mean filter deleted

•  Reason: Data with low resolution, lose edge information when using mean filtering

•  Thresholding method changed •  From Balanced Histogram Thresholding (BHT) to local

thresholding •  Reason:

•  Low contrast between hippocampus and neighboring regions •  Non-uniform intensity for interested regions

•  Erosion-Region growing added •  Reason: gain a more accurate boundary

Image

local block

local mean

Local thresholding

Results-parameters

• Optimal Parameters •  For Dataset C1(1.5T):

•  For Dataset C2(3.0T):

w1 w2 P1 P2

3 20 (w/3,h/2-20) (2w/3,h/2+20)

filter size w1 block control points P1,P2

Min/max filtering Thresholding Middle-block

selection Largest

Connected Component

Erosion-Region growing

Local block size w2

w: image width, h: image height

w1 w2 P1 P2

4 23 (w/3,h/3) (2w/3,h/3+60)

Results-evaluation

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

HF009 HF010 HF011 HF012 HF013 HF014 HF015

Dice Coefficient for C1(1.5T)

Dice Coefficient

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

HF021 HF022 HF023 HF024 HF025

Dice Coefficient for C2(3.0T)

Dice Coefficient

Mean dice coefficient : 0.6 Mean dice coefficient : 0.672

Discussion • Poor outcomes:

•  HF015

• Good outcomes: •  HF009

Dice Coefficient: 0.23

ground truth our result

ground truth our result

Dice Coefficient: 0.78

1.  Local variations inside hippocampus

2.  Errors caused by transferring from coronal to sagittal view

5/3/16

4

Summary • Min/max filtering and local filtering can correct non-uniform backgrounds and enhance boundaries

•  Parameters for MR imaging highly affect the performance of image analysis algorithms

• Our method is sensitive to middle-block selection

• Our result didn’t meet our expectation but close

picture from: www.timothy-carter.com