Prezentacja multimedialna współfinansowana przez Unię Europejską w ramach Europejskiego Funduszu Społecznego w projekcie Andrzej Materka Medical Electronics: Imaging (2) Innowacyjna dydaktyka bez ograniczeń zintegrowany rozwój Politechniki Łódzkiej zarządzanie Uczelnią, nowoczesna oferta edukacyjna i wzmacniania zdolności do zatrudniania osób niepełnosprawnych Zadanie nr 27 Nowy kierunek nauczania w języku angielskim - Biomedical Engineering - studia stacjonarne I stopnia 90-924 Łódź, ul. Żeromskiego 116, tel. 042 631 28 83 www.kapitalludzki.p.lodz.pl
Model of image acquisition system Image formation system Digital image (2D, 2½D, 3D, video) Investigated object 2
Examples of image acquisition methods Source of energy Acquisition technique Excitation Carrier of the visualized information Radiography X-ray radiation X-ray radiation External Internal Mammography As above As above Tomografia komputerowa As above As above Ultrasonography Acoustic wave Acoustic wave Visual imaging Visual light Visual light Scintigraphy, SPECT PET Gamma radiation of isotope marker Gamma photons emitted at positon anihilation (decay of isotope marker) Thermography Infrared radiation Internal and external Magnetic resonance imaging Pulse of magnetic field or radio-frequency waveform Radiowave Fluorescence microscopy Laser light Visual light (fluorescence) 3
Image formation system: X-ray examination Subject body Source of X-ray radiation Bone Film or detector Fluorescent screen 4
Image formation system: computed tomograhy Source of X-ray radiation Subject body Rotation Array of detectors Image reconstruction 5
Theory and techniques of medical imaging applications Application (diagnosis, therapy) Physiology of investigated organ Physics of imaging Image acquisition instrument (scanner) Image processing, analysis, and numerical modeling Need for collaboration of engineers and scientists of different disciplines! 6
Diagnostic system design Physiology of investigated organ Physics of imaging Properties of objects (tissues, organs,...) (physiology, static/moving, density of matter, anatomy/function) Physical phenomena of image formation (source of exciting energy, measured signal) Properties of instrumentation (signal to noise ratio, linearity and efficiency of detectors, artefacts) Method of image formation (scanning trajectory, MRI sequence protocol, timefrequency resolution compromise, method of image reconstruction) Methods of image processing and analysis Application (diagnosis, therapy) Image processing, analysis, and numerical modeling Image acquisition instrument (scanner) 7
Statistical approach to image analysis Measurement uncertainty errors Biological variability (tissues, organs, patients, diseases, normal state, pathology) Image acquisition technique (resolution, image discretization, object projection, intensity quantization, geometric distortion, body movement during measurement, detector noise) Image processing methods (parameters og image preprocessing filter, binarization threshold, image compression level) 8
Evaluation of image acquisition techniques Image analysis always involves a randomness. Table of diagnostic test results Person diagnosed as healthy Person diagnosed as non-healthy TN: true negative, FP: false positive Healthy person TN FP Non-healthy person FN TP FN: false negative TP: true positive Sensitivity Specificity Example n = 300 persons, n1 = 100 healthy, n2 = 200 non-healthy TP = 97 persons, FN = 3 persons TN = 176 persons FP = 24 persons C = 0.97, S = 0.88 9
Diagnostic system design Sensitivity Specificity Maximization of sensitivity (test of disease occurence is positive when a person is ill indeed) Maximization of specificity (test does not give a positive results when a person is healthy) 10
Goals and stages of image analysis Object parameters measurement using image segmentation Object parameters measurement through model fitting Automatic image interpretation 11
Object parameters measurement using image segmentation 3D scene Image acquisition Preprocessing Thresholding Postprocessing Image segmentation Feature extraction 3D scene parameters 12
Preprocessing example: image averaging Assumptions x k [..] x[i,j] = x*[i,j] + v[i,j] y[..] true intensity noise - the mean value of noise is zero - samples of noise are statistically independent and have the same probability density distribution - the visualized object does not change its position K number of averaged images Mean value: μ=x*[i,j] Variance: σ 2 /K 13
Preprocessing example: nonuniform illumination Grayscale image Axonometric projection Histogram Binary image 14
Preprocessing example: correction of nonuniform illumination Grayscale image Axonometric projection Histogram Binary image 15
Example: quantitative analysis of axons cross-sections http://pl.wikipedia.org/wiki/kolimator Image acquisition - tissue sample, - illumination, - microscope, - digital camera, - identification of geometric distortion, - calibration. 16
Example: quantitative analysis of axons cross-sections Preprocessing - correction of geometric distortion, - reduction of average brightness nonuniformity. Thresholding Image histogram 17
Example: quantitative analysis of axons cross-sections Postprocessing - removal of regions touching image boundary, - filling of holes, - contour smoothing, - removal of small regions. 18
Example: quantitative analysis of axons cross-sections A P F x F y # [µm 2 ] [µm] [µm] [µm] 1 858 117 38 36 2 2063 179 61 51 3 649 110 23 45 4 1034 133 37 46 5 405 81 24 29 6 1353 144 45 47 7 928 118 34 39 8 938 139 37 49 9 497 102 19 42 10 951 119 38 36 11 1298 148 38 51 12 1215 142 36 51 13 195 57 16 20 14 1489 227 66 58 Feature extraction - object identification and labelling, - finding contours, - calculation of geometric parameters (features), - fitting models. 15 974 127 41 37 19
Example: measuring width of a stripe Image of a stripe Axonometric projection w 20
Example: measuring width of a stripe Binary image after thresholding 21
Example: measuring width of a stripe Object (bright stripe on dark background) Object brightness profile (along line AB) Analog image brightness profile Analog image brightness profile (noise added) 22
Example: measuring width of a stripe - discretization error points of bright stripe points of dark background : uniform distribution, between and For each the result is the same! Estimator of stripe width Standard deviation of the estimator The estimator is unbiased, but is not consistent. 23
Errors caused by segmentation Local threshold Global threshold Thresholding (constant threshold value (global threshold), at nonuniform brightness of background and objects, changes the shape of objects) Discretization (details of analog image cannot be distinguished within pixels) Binarization (all the gray levels are replaced by one of two values of brighntess) Segmentation introduces irreversible loss of information contained in the image. 24
Object parameters measurement through model fitting 3D scene Image acquisition Preprocessing Model fitting 3D scene parameters 25
Example: measuring width of a stripe Image of a stripe Axonometric projection w Model - brightness profile along image line 26
Model example: parameterized brightness profile 27
Model of brightness variation at the edge of the stripe Δ V V B i x L x R 28
Example: model fitting to edges of the stripe Stripe width estimator Subpixel accuracy 29
Model fitting: active contour (snake) Active contour Node Analysed object dr P. Szczypiński - internal forces of the contour curve - image interaction component, - external forces, 30
3D model of heart muscle for visualization Heart muscle evolution MRI cross-sections Blood flow dr P. Makowski 31
3D model of heart muscle for visualization Triangulation of cross-sections Holobench 3D visualization workstation Grid model, 3D dr P. Makowski Reconstructed heart walls 32
3D model of heart muscle for visualization dr P. Makowski 33
Automatic image interpretation 3D scene image understanding computer vision Image acquisition Preprocessing Model fitting Segmentation Model fitting Feature extraction Model fitting Pattern recognition 3D scene parameters supervised methods unsupervised methods 34
Example: automatic texture segmentation Texture mosaic Image after segmentation http://www-dbv.informatik.uni-bonn.de/image 35
Pattern recognition Grass, class ω 1 Bark, class ω 2 36
Pattern recognition Grass, class ω 1 Bark, class ω 2 sample Texture features (for each sample) y1=s(2,0)sumentrp y2=s(1,1)sumentrp 64 feature vectors: 32 for class ω 1 + 32 for class ω 2 37
Pattern recognition D(Y) 64 points in the feature space: 32 for ω 1 + 32 for ω 2 38
Pattern recognition Pattern: Y ω ω Ω, Ω={ω 1, ω 2,, ω K } Vector of features (observations) Class indicator A set of K classes [y 1, y 2,...y p ] 1 ω 1 [y 1, y 2,...y p ] 2 ω 1 [y 1, y 2,...y p ] N/2+1 ω 2 [y 1, y 2,...y p ] N/2+2 ω 2 [y 1, y 2,...y p ] N/2 ω 1 [y 1, y 2,...y p ] N ω 2 ω 1 ω 2 Classifier ω k 39
Pattern recognition Discriminant functions: D k (Y), k=1,2,,k If Y~ω j, then The border between classes ω j and ω k : D j (Y)-D k (Y) = 0 (hipersurface) Example: D(Y)=D 1 (Y)-D 2 (Y) = 0 (a straight line) D(Y) 40
Pattern recognition: supervised training of a classifier - training set Tuning of weigths - overfitting - test set Observation Y Classifier ω j Class predicted by the classifier Training algorithm Known class Error calculation ω j Example: Linear classifier 41
Pattern recognition: brain MRI example Scull + test objects Regions of interest 42
Pattern recognition: brain MRI example Polystyren beads of different size Beads suspended in agar gel (test objects) MRI cross-section of test objects (2-3.2 mm, 0.8-1.3 mm) D. Jirak, M. Dezortova, M. Hajek, Praha 43
Pattern recognition: brain MRI example Clusters in texture features space 44
Information fusion, image registration http://visiblehuman.epfl.ch/ Head slice CT MRI 45