Can Color Detect Cancer?

1
Can Color Detect Cancer?

• Andrew Rabinovich
• 12/5/02

2
Dead or Not?
E 300 cancerous ? DEAD
F 0 cancerous ? HEALTHY
3
How To Detect Cancer?

• Spectral Information
• Spetial Information ? Texture

4
Spectral Information Analysis

• Proper Image Acquisition
• Pre-processing(image registration)
• Color Information Extraction

5
Image Acquisition
RGB vs. Hyperspectral
6
Image Registration

• Registering spectral bands with each other
• is absolutely unavoidable!!!
• Acquisition system instability optical
• aberrations result in spectral stack
• misalignment

7
Raw Spectral Data
Short Band Pass (Blue)
Long Band Pass (Red)
8
Misalignment
9
Misalignment
10
Registration of Multi modal Images

• No brightness constancy
• Common features at high resolution
• Individual features at low resolution
• Suppress the individual and extract the common
  using a high pass filter

11
Laplacian of Gaussian Filter
0.1 0.5 1
5 (-1.9694, 2.1693) (-1.7186, 2.0336) (-1.9646, 2.1624)
10 (-1.9264, 2.1329) (-1.8773, 2.1047) -1.9599 2.1592
20 (-1.8815, 2.1150) (-1.7773, 2.0511) -1.9559 2.1633
50 (-1.8809, 2.1283) (-1.7986, 2.0602) -1.9472 2.1762
Mean Shift (-1.8970, 2.1253) Mean Shift (-1.8970, 2.1253) Mean Shift (-1.8970, 2.1253) Mean Shift (-1.8970, 2.1253)
12
Filtered Images
Low Band Filtered
High Band Filtered
13
Shi Tomasi Affine Registration
Determine the motion based on an Affine
transformation
Transformation is found to sub-pixel resolution
14
Registered Spectral Images
15
Registered Spectral Images
16
Before and After
17
Color Models to Extract Spectral Signal

• Color Deconvolution
• Non-Negative Matrix Factorization
• Independent Components Analysis

18
Color Deconvolution
19
Non-Negative Matrix Factorization
20
ICA
21
Discussion

• To quantify the separation of spectral signals,
  each of the dies must be imaged independently and
  compared with the separated signal
• This study was done with RGB, however,
  Hyperspectral is a MUST