1. Near Real Time Early
Esophageal Cancer
Detection Using a
Graphics Processing
Unit
Jason Helms
2. Introduction
Purpose
Image Processing
Discrete Wavelet Transform
Fractal Dimension Decomposition
General Methodology
Phase Breakdown
Base Results
Threshold Identification
Final Results
Timing Results
Future Work
Conclusions
Agenda
3. Introduction
Jason Helms
BS in CS at EWU
Medical processing has been recurrence
“Grand Challenges in Biomedical Image Analysis”
“Sub-Challenge: Early Barrett’s cancer detection”
4. Purpose
Identify Lesions
Input: Endoscopic image
Current: One image at a time
Extended future goal: Video application
Why: Getting data expensive, reduce patient
burden, optimize input data gathering
Near Real Time
Saga University time: 3 minutes 0 seconds
Problems: Few unknowns in regards to
aspects of the paper.
Goal: < 10s/image
Why: 0:30 video @ 30 fps = 900 images
9000 s = 2.5 hours.
8. [ 2 ]
Purpose
J. Yamaguchi, A. Yoneyama, and T. Minamoto, “Automatic detection of early
esophageal cancer from endoscope image using fractal dimension and discrete
wavelet transform,”
Original Work
Published 2015
Base method
12. OpenCV
- GPU integration in 2011
- Version used: 3.0.0
- OpenCV Focus: Real Time
Image Processing
- Diverse Hardware Compatibility
[ 4 ]
13. Qt Creator: used to manage phases
C++: OpenCV, QT Creator
C++ & QT Creator
[ 5 ]
14. Used for:
- Input reduction
- Small Vein Filter
- Minimal loss of input details
The first DWT was invented by the
Hungarian mathematician Alfréd
Haar.
Discrete Wavelet
Transform
22. First convert image to Binary using
Dynamic Thresholding.
Recursively cover image in grid of
boxes
Count boxes with a 1
Compute: log(# of boxes with one’
s) divided by log (length of box
side) and save in variable (DT)
Box Counting
24. Box Counting
Final Steps
After all boxes have been
processed the following calculated
for each color:
Call this D{C} where C is from {R,G,
B,L} then:
Total Box Result (BT) for a block is:
BT = DR * DB * DG * DL
26. General Methodology
Phase 1
Codename: clipping
Purpose: preprocessing
Difficulty: Eh?
Phase 2
Codename: pre-wavelet
Purpose: preprocessing
Difficulty: Easy
Phase 3
Codename: DWT
Purpose: final preprocessing
Difficulty: Medium
Phase 4
Codename: Box Count
Purpose: detect patterns
Difficult: Hard
27. Phase 1
RGB Decomp
Input is a PNG
3 Channel
RGB
Lum. Calc.
No standard
Y = a * R + b * G + c * B
Size Reduce
Input Dimensions: 1600 x 1200
Need: 1024 x 1024
3 channel -> single channel
Edge Enhance.
Gaussian Fuzzy Edge
Enhancement
Brightness Enhance
28. Sample input image, not test image
Red, Green, Blue and Luminance
decomposition
Next up
39. DWT reduction sample
Small Vein Filter - convert & magnify
Small Vein Filter - magnify and dwt 1
Small Vein Filter - dwt 2
Small Vein Filter - before after
Next up
46. Phase 4
Bin. Image
Input: single channel (grayscale)
image
Output: Binary image
Box Counting
Input: 16 x 16 pixel binary image
Output: float
Results?
Still need statistics, but we can
see something, too many to be
coincidence
High false positive rate
Grid Reduction
Input: 32 x 32
Output: four 16x16 pixel images
Non-overlapping
47. Phase 4 set up
Color Intensity to Binary
Box count output
Next up
56. From one promising square to
identifying all of the suspected area
marked by physicians.
Begins with color intensity
thresholding
Threshold
Identification
59. Converting an image to binary
uncovered yet another threshold set
to a blind average: Dynamic
thresholding
Relaxed this threshold
Threshold
Identification
61. Identified another in DWT, left at
blind average
Did not make sense (visually) to
change
However, Daubechies with D=4
would be better but costly
Threshold
Identification
83. Note*
All software libraries open source
All IDE’s, text editors and pdf creators open source
Open source inline with Grand Challenges community
Open source important to technological advancements
84. References
[ 1 ] S. K. Bram van Ginneken and C. Shneider., 2012-2016. Available: http:
//endovissub-barrett.grand-challenge.org/
[ 2 ] S. K. Bram van Ginneken and C. Shneider., 2012-2016. Available: http:
//grand-challenge.org/
[ 3 ] http://www.clipartbest.com/clipart-bTyoR8Xrc
[ 4 ] "Lenna 97: A Complete Story of Lenna". Dr. Lai Man P web pages. City
University of Hong Kong. 28 July 1997. Archived from the original on 2012-12-04.
Retrieved 24 September 2013.
86. References
[ 8 ] D. Bryant. (2013, mar) Fourier transforms. [Online]. Available: http:
//devonbryant.github.io/images/fourier/fourier transform.png
[ 9 ] J. Yamaguchi, A. Yoneyama, and T. Minamoto, “Automatic detection of early
esophageal cancer from endoscope image using fractal dimension and discrete
wavelet transform,” in Information Technology - New Generations (ITNG), 2015
12th International Conference on, April 2015, pp. 317–322.
[ 10 ] D. Pearcy. (2013, may) Introduction to fractal geometry. [Online]. Available:
https://danpearcymaths.files.wordpress.com/2013/05/image21.png