Hierarchical Image Segmentation for Identifying Stroke Regions In Apparent Diffusion Coefficient (ADC) Image Maps Anthony Bianchi BRITE @ UCR 2007 Advisor:

Hierarchical Image Segmentation for Identifying Stroke Regions

In Apparent Diffusion Coefficient (ADC) Image Maps

Anthony BianchiBRITE @ UCR 2007Advisor: Bir Bhanu

8/24/2007

Outline

• Background• How it Happens• The Images• Why Automatic is Needed

• Process• Flowchart• Automatic Thresholding• Connected Components• Example

• Results• Two Patients• Findings

• Conclusion• Acknowledgments

• Stroke: 1/4000 live births• Arterial-ischemic stroke, 73%

Arterial Ischemic Stroke

Cerebral arterial thrombosis: possible postnatal etiology of

AIS.AIS AIS Focal Lesion Focal Lesion

Background

Apparent Diffusion Coefficient (ADC) Image Maps≤ 5 days

≥ 5 days

25 days

5 days

• An ADC image map measures the diffusion of water. If the diffusion is low the grayscale value is low.

Why is Automatic Segmentation Needed?

• Currently manual segmentations is time intensive and inaccurate. Manual segmentations can very over 30% from one person to the next, and can take hours per patient.

• An automatic segmentation algorithm will be repeatable, and will take minutes per patient.

• We are currently working with LLUMC. They would like to use this segmentation to classify stroke victims into mild, moderate, and severe. They will use these labels to accept patients for stem cell trials.

Find Image to be Segmented

Find Threshold Automatically

Split Image Using Threshold

Find Connected Components That Satisfy

Second threshold

Lower Region

Higher Region

Close and Fill Images

# Regions > 1YES YES

NO NO

Find Connected Components That Satisfy

Second threshold

# Regions > 1

Separation Results

The Process

Automatic Thresholding: Otsu’s method

Threshold Found 176

2w (t) = q1(t) 2

1(t) + q2(t) 22(t)

2w (t) = q1(t) 2

1(t) + q2(t) 22(t)

Within Group Variance

Sum of Probability in

Group 1

Variance Group 1

Sum of Probability in

Group 2

Variance Group 2

• We test every threshold to find the smallest Within Group Variance.

• A recursive form of the above equation is implemented to cut down computation time.

Connected Components

1 1 11 x 00 0 0

Mask

1 1 0 1 1 1 0 11 1 0 1 0 1 0 11 1 1 1 0 0 0 10 0 0 0 0 0 0 11 1 1 1 0 1 0 10 0 0 1 0 1 0 11 1 0 1 0 0 0 11 1 0 1 0 1 1 1

1 1 0 1 1 1 0 21 1 0 1 0 1 0 21 1 1 1 0 0 0 20 0 0 0 0 0 0 23 3 3 3 0 4 0 20 0 0 3 0 4 0 25 5 0 3 0 0 0 25 5 0 3 0 2 2 2

Example of Connected Components

• A mask gets sent through the image. Each pixels is evaluated by the mask to see if it has a neighboring pixel. If there is a neighboring pixel the selected pixel gains the same label of that pixel. If no neighboring pixel is found a new label is created for that pixel.

Example of the ProcessFind object to be

segmented.

Threshold found 149

Object > 50% found

Threshold found 98

Object found closed and filled

Patient 1 Patient 2RED = Damage GREEN = Area RED = Damage GREEN = Area

SliceTotal Area

(pixels)Damaged

Area (pixels)Percent

Damaged6 13405 149 1.11%7 13340 1501 11.25%8 13152 2686 20.42%9 12580 3110 24.72%10 12057 2794 23.17%11 11638 2496 21.45%12 9448 1394 14.75%13 11056 2078 18.80%14 7499 307 4.09%15 5410 0 0.00%16 3036 78 2.57%17 349 0 0.00%

Total 112970 16593 14.69%

SliceTotal Area

(pixels)Damaged

Area (pixels)Percent

Damaged1 4452 0 0.00%2 5381 0 0.00%3 6444 0 0.00%4 7379 79 1.07%5 8540 226 2.65%6 9614 794 8.26%7 10594 707 6.67%8 11311 583 5.15%9 12133 552 4.55%

10 12501 608 4.86%11 12141 275 2.27%12 11623 141 1.21%

Total 112113 3965 3.54%

• For patient 1 the automatically segmented data gave a total damage of 14.7%, while manually segmented images gave a total of 17.7% damage.

•The reason for the difference between the manual and automatic segmented is because the area used in finding total area in the automatic segmented included spinal fluid. This fluid can be found by the automatic method and can be removed.

•For patient 2 we found the damaged area to be 3.5%, and the manual segmentation gave a 3.6% result. For patient two the manual segmentation included the cerebral spinal fluid, which was included in the area.

Results

Conclusion

• The experiments show that the automatic method has small differences compared to the manually segmented images. But, it is effective and consistent in finding the damage area in the ADC images.

• Hypoxic-Ischemic Encephalopathy is another type of stroke that happens every 1/1000 live births. These injuries are diffused through the brain unlike the AIS patients. This segmentation method should be able to detect this type of stroke.

• The next step is trying to use this method on different MRI types such as T2 image maps.

• A 3D approach could give better results, because it could connect the structure from slide to slide.

Acknowledgments

• I would like to thank:

• Dr. Bhanu for his guidance.

• Jacqueline Coats, Andy Obenaus, and Stephen Ashwal (from LLUMC) for data and useful information.

• BRITE advisors for the opportunity.

• Friends & Family for Support.