Transcript Biomedical Image Analysis Rangaraj M. Rangayyan
Biomedical Image Analysis
Rangaraj M. Rangayyan
Ch. 5 Detection of Regions of Interest: Sections 5.4-5.7, 5.10-5.11
Presentation March 3rd 2005 Jukka Parviainen Yevhen Hlushchuk
2 Outline
segmentation – an ideal example problems in biomedical context categories for segmentation methods two methods explained with more details – detection of pectoral muscle in mammograms using Hough transform (section 5.10.1) summary & discussion T-61.182 Parviainen, Hlushchuk March 3rd 2005
3 Books at table
Sonka, Hlavac, Boyle: ”Image processing, analysis and machine vision” – chapter 5: Segmentation – similar terminology, examples from Rangayyan Gonzalez, Woods: ”Digital image processing” – chapter 9: Morphological image processing – – chapter 10: Image segmentation introduction: lots of biomedical applications Rangayyan includes some advanced methods T-61.182 Parviainen, Hlushchuk March 3rd 2005
4 What is region of interest (ROI)?
divide image into regions that correspond to structural units examples in mammograms: – tumors and masses – – pectoral muscle calcifications ROIs are detected using properties of – discontinuity = edges – similarity = regions T-61.182 Parviainen, Hlushchuk March 3rd 2005
5 What is process ”segmentation”?
segmentation reduces pixel data to region based information highly application dependent simpliest case: thresholding gray-scale pixel values (Fig. 5.1) T-61.182 Parviainen, Hlushchuk March 3rd 2005
6 Practical problems which make it a tough job!
noise, noise, noise – derivatives are sensitive to noise, LoG especially low dynamic range – no exact borders in images (Fig. 5.32a, etc) stochastic algorithms: – need for a proper seed for region growing T-61.182 Parviainen, Hlushchuk March 3rd 2005
Categories for segmentation methods 7
T-61.182 Parviainen, Hlushchuk March 3rd 2005
8 Categories for segmentation methods
thresholding (M1) – problem: global, neglates all spatial information boundary-based (M2) – problem: edge segments to boundaries region-based (M3) – problem: selection of homogeneity criterion T-61.182 Parviainen, Hlushchuk March 3rd 2005
9 M1 Thresholding
class: all pixels whose values within a certain range determined by valleys in the image histogram – background and objects not always having bimodal histogram (Fig. 5.4/Sonka) optimal thresholding may fail due to illumination T-61.182 Parviainen, Hlushchuk March 3rd 2005
10 M2 Boundary-based methods
disjoint edge segments to closed-loop boundaries is a difficult job edge detection using gradient masks – – gradient magnitude and direction edge-flow propagation (p. 493) global
Hough transform
(section 5.6) T-61.182 Parviainen, Hlushchuk March 3rd 2005
11 M3 Region-based methods
region growing –
pixel aggregation using additive tolerance
/ multiplicative tolerance region splitting/merging – – split region into a non-overlapping set of subregions which all fulfill conditions or predicates P usually quadtrees – adjacent similar subregions can be merged T-61.182 Parviainen, Hlushchuk March 3rd 2005
12 M4 Other advanced methods and techniques
morphological watershed fuzzy-set-based region growing (section 5.5) – fuzzy membership, crisp boundaries linear prediction for proper seeds (section 5.4.10) improvement of contour or region estimates (section 5.7) T-61.182 Parviainen, Hlushchuk March 3rd 2005
Method #1: Region growing using an additive tolerance 13
T-61.182 Parviainen, Hlushchuk March 3rd 2005
14 Pixel aggregation using additive tolerance (section 5.4.4)
compare properties of spatially neighboring pixels with those of seed pixel (Fig. 5.17) add pixel f(m,n) if |f(m,n)-seed| <= T what is a good seed?
add pixel f(m,n) if |f(m,n)-mu_R| <= T where mu_R running mean...
T-61.182 Parviainen, Hlushchuk March 3rd 2005
Method #2: Hough transform 15
T-61.182 Parviainen, Hlushchuk March 3rd 2005
Detection of objects of known geometry – Hough transform 16
objects in images may sometimes be represented in an analytical form, such as straight-lines, circles, ellipses, parabolas Hough transform converts images to parametric plane, where analytical forms may be found easier (section 5.6) study of Hough transform and Gabor wavelet based methods in mammogram data (5.10) T-61.182 Parviainen, Hlushchuk March 3rd 2005
17 Hough: mapping to parameter space
points at straight line yi = k xi + b, where k is slope and b (Fig 10.17/G) k and b are not limited – use rho and theta instead now each point corresponds a sinusoidal – line in original figure can be found as intersection of curves T-61.182 Parviainen, Hlushchuk March 3rd 2005
18 Hough: example with 5 points
five labeled points {1,...,5} (Fig. 10.20/G) top-right: five sinusoidals in parameter space bottom-left: [A] intersection of curves corresponding {1,3,5} at rho=0, theta = -45 deg; [B] similarly rho=0.707D, theta = +45 deg edge linking: compute gradient; subdivide rho and theta into bins; count T-61.182 Parviainen, Hlushchuk March 3rd 2005
19 Application: Detection of pectoral muscle in mammograms (5.10.1)
identify points {N1,..., N6} and ROI N1-N2-N3-N4 (Fig 5.64) geometric and anatomical constraints – p. muscle theta = {120 .. 170} deg, intersects N1-N2, ...
LP + Sobel gradients in ROI count Hough accumulator cells eliminate impossible lines choose most likely (max) line T-61.182 Parviainen, Hlushchuk March 3rd 2005
Application: Detection of pectoral muscle in mammograms 2
result:
20
T-61.182 Parviainen, Hlushchuk March 3rd 2005
21 Summary & discussion
computer analysis starts with segmentation regions of interest (ROI) highly application dependent methods several large studies in the book comparing different segmentation methods T-61.182 Parviainen, Hlushchuk March 3rd 2005
22 Matlab Image Processing Toolbox
help images version Matlab 5.3 - 7, IPT 2.2 - 4 – roidemo (enhancement) – qtdemo* (quadtree), edgedemo* – help iptdemos T-61.182 Parviainen, Hlushchuk March 3rd 2005