FAPBEDCheckpoint Presentation:

Feature Identification

Danilo ScepanovicJosh Kirshtein

Mentor: Ameet Jain

Sample Image

Difficult Surface

To Detect

•Faint Edges

•Edges In Close Proximity

•Relevance To Larger Problem Of Segmentation

Identified Properties• Pixel Density Value• Linear Gradient• Maximum 2D Gradient and Directionality• Pixel Disparity Magnification / Intensification

More Properties to Analyze• Principle Component Analysis• Weighted Incidence Angles

Methods

• Linear Gradient

• Thresholding

• Close Proximity Edge Enhancement

• 2D Gradient

• Intensification

Linear Gradient

• Look at gradients along X and Y direction independently

• Detect edges by observing:– Raw pixel values– Gradient values along single axis– Range of gradient values along single axis

• Future: Weight by normal to surface as detected by 2D gradient analysis

Y = 285

Raw Pixel Value

Gradient Value

Range of Gradient

Raw Pixel Value

Gradient Value

Range of Gradient

X = 215

Thresholding• Densities are systematically

distributed within a slice and a volume

• Thresholding separates main classesPixel Densities from Original

Slices

Derivative of

Pixel Densities

Play Threshold Movie

• Notice loss of soft tissue occurs between 50-70

• Insides of bones disappear between 70-80

• Above that, bone edges disapear

Thresholding Characteristics

Close Proximity Edge Enhancer

• Apply a filter that will enhance gaps between bones in close proximity

• Involves looking at some number of neighbors and adjusting pixel values

• Good at reducing pixel values that lie between bones (max pixel values unchanged)

• Future: Use to enhance detection at bone junctions

How do we get more information

from the image?

2D Gradient

• Convolve image with 2D gradient detector:– Maximal gradient– Direction of max gradient

• Results: Enhances all edges in image

• Future: Use to enhance confidence in a detected edge and to perform PCA and/or Weighted Incidence Angle analysis

First 2D Gradient Filter

• Compute gradient across entire diameter of box (8 directions)

• Pick max value

• Determine direction

Window Size = 3

Play Edge Movie

Window Size = 3

Window Size = 5

Window Size = 7

Arrows Indicate Direction of

Maximum Gradient

Second 2D Gradient Filter

• Compute gradient originating from center of box (8 directions)

• Pick max value

• Determine direction

Window Size = 5

Window Size = 3

Window Size = 5

Window Size = 7

Comparison of both methods

Method 1 (Window = 3)

Method 2 (Window = 3)

Difference

Intensifier

• Increase pixel densities that lie above the local mean

• Decrease pixel densities that lie below the local mean

Play Intensifier Movies

Intensifier Movies1) As average box size increases, edges

become thicker while soft tissue noise is suppressed

2) Smaller box size correlates with larger speckle and image obfuscation

– Optimal clarity is achieved after first few feedback-loop iterations

– Forcing hard classification introduces significant noise and results in information loss

3) Increasing box size yields thicker edges4) Compounding final images from different

box sizes yields more information

Timeline

Item Target Date StatusBackground Reading 27-Feb CompleteThresholding Algorithm Implementation 12-Mar CompleteNeural Network Attempted 26-Mar CompleteSpeed Ups 16-Apr In ProgressFinal Program Evaluation 23-Apr In ProgressFine Tune 30-Apr Awaits

Hurdles

• Difficulties– Finding properties of surfaces– Combining different results into coherent image– Starting to implement methods

• Dependencies Not Met– None

• Thanks to:Ameet Jain

Ofri Sadowski

Dr Russell Taylor

Mathworks