Green Algae Cell Identification Devin Karns

advertisement
Green Algae Cell Identification
Devin Karns
Project
• Find and segment cells in microscopy videos
• Analyzed videos of two kinds of cells
▫ Chlamydomonas cells
▫ Red blood cells
Tasks
• MATLAB code is broken up into four main
sections
▫
▫
▫
▫
Preliminary first frame measurements
Frame by frame analysis
Accumulated centroid processing
Video assembly
Preliminary (part 1/3)
• Opens video file and determines number of
frames
• Converts to grayscale and performs adaptive
histogram equalization
• Performs averaging and opening subtraction
Preliminary (part 1/3)
Preliminary (part 1/3)
Preliminary (part 2/3)
• Entropy-based edge detection (El-Sayed, 2011)
Preliminary (part 3/3)
• Find mean radius from circular hough transform
Frame by Frame (part 1/6)
• Remove further noise with scaled structure
element and get edges
Frame by Frame (part 1/6)
Frame by Frame (part 2/6)
• Compute watershed
• Filter regions by area
• Check each region with its described ellipse from
regionprops
Frame by Frame (part 2/6)
Frame by Frame ( part 2/6)
Frame by Frame (part 3/6)
• Cross check filtered
regions with
watershedprocessed edge
image regions
• Check ellipse fit of
each region
Frame by Frame (part 3/6)
Frame by Frame (part 3/6)
Frame by Frame (part 4/6)
• Find rejected regions
Frame by Frame (part 4/6)
Frame by Frame (part 5/6)
• Get canny edge image
of filtered image
• Perform circular hough
transform on edge
image
• Mask filled hough
circles with rejected
region image
• Collect masked regions
• Iterate until threshold
is met
Frame by Frame (part 5/6)
Frame by Frame (part 5/6)
Frame by Frame (part 6/6)
• Check ellipse fit of recovered region
combinations and accumulate
Centroid Processing
• Generate 3d centroid
position array
• Perform 3d convolution
with spherical structure
element
• Threshold and mask
centroid positions to
eliminate transient
artifacts
Centroid Processing
Centroid Processing
Assemble Video
• Highlight found regions
• Circle found centroids
Assembled Videos
• Media player doesn’t want to play the videos.
Lets see if I can play it somewhere else.
Problem
• Not all cells are identified
• Method is susceptible to failure when cells
clump, overlap, or pass by the edge of the image
• False positives occur sporadically in empty
regions with borders of high entropy
• Optimal thresholding may provide better results
• Videos with objects of many sizes tend to fail
• Very susceptible to noise
• Doesn’t work for videos with many internal
segments
Conclusion
• Many improvements are possible (better radii
guessing code, between frame learning, more
sophisticated region filtering criteria…)
• My approach is decent, far from perfect
• Skyfall was a bad movie
Questions
• Seriously, why do we have to have this scene?
Download