Generated Micro-Pattern Imagery

Transcription

1 Generated Micro-Pattern Imagery SSB Jason M. Kinser, D.Sc. jkinser@gmu.edu Physics, Astronomy and Computational Sciences George Mason Univ. February / 47

2 Outline 4 Gabor Filtering 1 Problem Statement 2 Previous Research 3 Approach 5 Pulse Images 6 Advertisement 2 / 47

3 Problem Statement Outline 1 Problem Statement 2 Previous Research 3 Approach 4 Gabor Filtering 5 Pulse Images PCNN ICM 6 Advertisement 3 / 47

4 Problem Statement Image Cells Actin Fibers Multiple types Drug reaction 4 / 47

5 Problem Statement Issues Different Regions Orientation Different shapes 5 / 47

6 Problem Statement Micro-Patterns 6 / 47

7 Problem Statement Advantages Predictable geometry Several geometries are possible No orientation issues 7 / 47

8 Problem Statement Different Patterns 8 / 47

9 Problem Statement Questions 1 Do drug dosages affect micro-pattern images? 2 Is it possible to generate realistic images? 3 Can we measure the similarity between generated and original images? 4 Is it possible to optimize image generation in a meaningful way? 9 / 47

10 Previous Research Outline 1 Problem Statement 2 Previous Research 3 Approach 4 Gabor Filtering 5 Pulse Images PCNN ICM 6 Advertisement 10 / 47

11 Previous Research Hamiltonians Head et al. define system: Start with random fibers. Cross-links Fibers are allowed to Bend and Stretch Optimize Hamiltoniian H T = µh S + κh B Not for Micro-Patterns Head, Levine, MacKintosh, Physical Review E, / 47

12 Previous Research Sliding Friction Walcott and Sun propose a system in which the fibers are like springs and have a sliding friction. The binding probability distribution n(x, t) is n t + ν n t = (1 N b) ρ a k d n where ν is the sliding rate, N b = ndx is the proportion of cross-links, ρ a is the attachment rate probability distribution, and k d is the detachment rate. Walcott and Sun, Proc. National Acad. of Sciences, / 47

13 Previous Research Texture Sharrif et al. create a model of actin fiber growth. To compare generated images to actual images they employ GLCM (Haralick texture measures). GLCM requires multiple quantization levels and directional matrices. 13 / 47

14 Approach Outline 1 Problem Statement 2 Previous Research 3 Approach 4 Gabor Filtering 5 Pulse Images PCNN ICM 6 Advertisement 14 / 47

15 Approach Concerns Micro-Patterns induce new constraints not considered by Head et al.. GLCM is insufficient and expensive. Models other than Head s to consider. Need metric that is sensitive to drug dosage. Multiple types of actin fibers. Cell is actually R / 47

16 Approach H S and H B 16 / 47

17 Approach H C and H F 17 / 47

18 Approach Theory H T = µh S + κh B + νh F + ξh C Four Energies to minimize: H S = stretch. H B = bend. H F = flow. H C = coherence. Cross-links from initial orientation - but can be broken. 18 / 47

19 Approach Multiple Types of Actin Fibers near perimeter are different than fibers in the interior. Growth fibers (perimeter) are thinner and move faster. 19 / 47

20 Approach Growth Fibers near perimeter are growing, branching and capping. Growth model considered but discarded because cells aren t growing much during the micro-pattern experiment. 20 / 47

21 Approach Third Dimension Cells are R 3. Fibers are along perimeter. Fibers on top of cell are not interacting with fibers on the bottom of the cell. Simulation is in R 2. So sets of fiber groups are established. Fibers do not have cross-links to fibers in other groups. 21 / 47

22 Approach Early Work 1 Blue: Interior (only one group) 2 Green: Upper perimeter fibers 3 Red: Lower perimeter fibers. Generated fibers over the original image. 22 / 47

23 Approach Optical Effects Original images are imaged through optical elements. Inherent diffraction due to finite size of elements. Other contents of cell besides fibers. 23 / 47

24 Approach Corrections Generated fibers. Point spread & Intensity Variations 24 / 47

25 Approach Movie Click for Fiber Movie 25 / 47

26 Approach Big Question How do we compare the original image to the generated image? 26 / 47

27 Gabor Filtering Outline 1 Problem Statement 2 Previous Research 3 Approach 4 Gabor Filtering 5 Pulse Images PCNN ICM 6 Advertisement 27 / 47

28 Gabor Filtering Theory g(x, y) = exp ( x 2 + γ 2 y 2 ) ) cos (2π x 2σ 2 λ + ψ, (1) 28 / 47

29 Gabor Filtering Correlations 29 / 47

30 Gabor Filtering Jets 1 Many filters = many correlations. 2 Creates data cube N V H. 3 Jet is an extraction. 30 / 47

31 Gabor Filtering Responses 31 / 47

32 Gabor Filtering Concerns 1 No sensitivity to F. 2 No sensitivity to ξ. 3 Large range between min and max. 32 / 47

33 Pulse Images Outline 1 Problem Statement 2 Previous Research 3 Approach 4 Gabor Filtering 5 Pulse Images PCNN ICM 6 Advertisement 33 / 47

34 Pulse Images PCNN Outline 1 Problem Statement 2 Previous Research 3 Approach 4 Gabor Filtering 5 Pulse Images PCNN ICM 6 Advertisement 34 / 47

35 Pulse Images PCNN Theory F ij [n] = e α F δ n F ij [n 1]+S ij +V F M ijkl Y kl [n 1] L ij [n] = e α Lδ n L ij [n 1] + V L W ijkl Y kl [n 1] kl kl U ij [n] = F ij [n] (1 + βl ij [n]) 35 / 47

36 Pulse Images PCNN Theory Y ij [n] = { 1 if U ij [n] > Θ ij [n 1] 0 Otherwise Θ ij [n] = e α Θδ n Θ ij [n 1] + V Θ Y ij [n] 36 / 47

37 Pulse Images PCNN Example 37 / 47

38 Pulse Images ICM Outline 1 Problem Statement 2 Previous Research 3 Approach 4 Gabor Filtering 5 Pulse Images PCNN ICM 6 Advertisement 38 / 47

39 Pulse Images ICM Theory Combines several models of mammalian visual cortex. F ij [n + 1] = ff ij [n] + S ij + W {Y[n]} ij { 1 if F ij [n + 1] > Θ ij [n] Y ij [n + 1] = 0 Otherwise Θ ij [n + 1] = gθ ij [n] + hy ij [n + 1] 39 / 47

40 Pulse Images ICM GMU - BINF - ICM 40 / 47

41 Pulse Images ICM Responses 41 / 47

42 Pulse Images ICM The Good News 1 Sensitivity to F. 2 Sensitivity to ξ. 3 Small range between min and max. 42 / 47

43 Pulse Images ICM Sensitivities Sensitivity to ξ and F. 43 / 47

44 Pulse Images ICM State of the Art 1 Sensitivity to ξ. 2 Possible to find best ξ for architecture. 3 Sensitivity to drug (preliminary success). 4 Possible to see difference in ξ w.r.t. drug dosage. 5 Recall ξ is parameter for H C. 6 Possible to understand mechanics of cell. 44 / 47

45 Pulse Images ICM State of the Future 1 PCNN vs. ICM 2 ν, µ, κ. 3 Relate back to biological knowledge of cells, actin and H parameters. 45 / 47

46 Advertisement Outline 1 Problem Statement 2 Previous Research 3 Approach 4 Gabor Filtering 5 Pulse Images PCNN ICM 6 Advertisement 46 / 47

47 Advertisement Third Edition 47 / 47