Extraction of Vessels from X-Ray Angiograms Titus Rosu Prof. Dr. Rupert Lasser Andreas Keil

Extraction of Vessels from X-Ray Angiograms

Titus Rosu

Prof. Dr. Rupert LasserAndreas Keil

Extraction of Vessels from X-Ray Angiograms 2

Introduction

Applications• Initial step for feature-based

algorithms– Intra-operative guidance– Reconstruction of coronary

arteries

Objective of the IDP• Segmentation of coronary

arteries• Implementing different

algorithms in the ITK framework• Testing and numerical evaluation

between the algorithms


Introduction

Data• (Rotational) angiography sequences

from stationary C-arms• Contrasted coronary arteries • No radial distortion • Pixel spacing is 0.3 x 0.3mm or 0.6 x 0.6mm

Problems• Projection images

Overlay of vessels and other structures• Varying contrast• Decreasing vessel width


Methods

Improving the images with image processing Algorithms (thresholding, normalizing, cropping...)

Segmentation Algorithms• Multiscale vessel enhancement filtering – (Frangi, 1998)• Multiscale detection of curvilinear structures in 2-D and 3-D image

data – (Koller, 1995)• Vessel segmentation using a shape driven flow – (Nain, 2004)

Segmentation• Assumption of a linear structure of the vessels• Eigen value analysis of the image intensities of every pixel• Analyzing with different scales (scale space)


Methods

Nain• Region based flow deforms the curve of interest• Using level set techniques to evolve the active contour• Determine the shape of a contour with a local ball filter (values

between 0-1)• small near-circle evolution• User input: Max. vessel width

Nain, Yezzi, and Turk. Vessel segmentation using a shape driven flow. MICCAI, vol. 3216 of LNCS, pp. 51-59. Springer, 2004


Methods

Frangi

• One of the standard papers on Hessian-based vessel filtering

• Basis for many other papers w.r.t– Usage of the Hessian– Multiscale analysis for vessels

Frangi, Niessen, Vincken, and Viergever. Multiscale vessel enhancement filtering. MICCAI, vol. 1496 of LNCS, pp. 130-137. Springer, 1998


Methods

Koller• Detect curvilinear structures of arbitrary shape • Using the 2nd derivation of the Gauß-Function to resolve the edges

left and right of the vessel-profile• Non linear algorithm => using positive min-function• Mulitscale analysis• User input min. and max. of the vessels width

Koller, Gerig, Székely, and Dettwiler. Multiscale detection of curvilinear structures in 2-D and 3-D image data. ICCV, pp. 864-869, 1995


Post processing and Segmentation

Frangi image post processing: • Double threshold => Segmentation



Koller image post processing: • centerline detection =>

=> double threshold =>=> connected line det. => segmentation

C. BLONDEL, G. MALANDAIN, R. VAILLANT, , and N. AYACHE, Reconstruction of Coronary Arteries From a Single Rotational X-Ray Projection Sequence, IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 25 (2006), pp. 653-663.



Koller image post processing: • double threshold => segmentation


Evaluation

• Quantitative comparison of results• Manually segmented data as “ground truth”• Same comparison conditions by post processing the result images

with different filters


Evaluation

• 3 images from different data sets for comparison• Comparison:

double threshold Frangi vs. double threshold Koller vs. ground truthdouble threshold Frangi vs. post processed centerline Koller vs. ground truth

• Numerical pixel wise comparison:


Evaluationmin. connected lines of 5 pixels Koller image

min. connected lines of 20 pixels Koller image

Double thres. Koller image

Double thres. Frangi image

Data set

ResultFN

ResultFP

ResultTP

ResultFN

ResultFP

ResultTP

ResultFN

ResultFP

ResultTP

ResultFN

ResultFP

ResultTP

71011C

0.57 0.36 0.43 0.60 0.27 0.40 0.53 0.42 0.47 0.55 0.41 0.45

70822E 0.59 0.21 0.41 0.61 0.16 0.39 0.55 0.20 0.45 0.52 0.17 0.48

70822B

0.59 0.37 0.41 0.63 0.26 0.37 0.56 0.33 0.44 0.54 0.30 0.46

best vs. worst


Evaluation

Merged double thres. Frangi, min. con. lines of 5 pixels Koller image

Merged double thres. Frangi, min. con. lines of 20 pixels Koller image

Merged double thres. Frangi, double thres. Koller image

Data set ResultFN

ResultFP

ResultTP

ResultFN

ResultFP

ResultTP

ResultFN

ResultFP

ResultTP

71011C 0.44 0.60 0.56 0.46 0.54 0.54 0.43 0.64 0.57

70822E 0.45 0.29 0.55 0.46 0.24 0.54 0.44 0.26 0.56

70822B 0.44 0.51 0.56 0.46 0.42 0.54 0.44 0.46 0.56

• Top: Frangi vs. Koller

• Bottom: Merged Frangi, Koller vs. manually seg. images

Overlapped pixels are segmented white, x = min(Frangi(x), Koller(x))

Merged Frangi, Koller images produce better TP/FN but increase FPs

best vs. worst


Evaluation

From top to bottom:

• Frangi vs. man. seg. images

• Min. connected component lines of 5 pixels Koller vs. man. seg. images

• Min. connected component lines of 20 pixels Koller vs. man. seg. images

• Double threshold Koller vs. man. seg. images

Overlapped pixels are segmented white, x = max(Frangi(x), Koller(x))


Evaluation

From top to bottom:

• Merged double threshold Frangi, min. connected component lines of 5 pixels Koller vs. man. seg. images

• Merged double threshold Frangi, min. connected component lines of 20 pixels Koller vs. man. seg. images

• Merged double threshold Frangi, Koller vs. man. seg. images

Overlapped pixels are segmented white, x = min(Frangi(x), Koller(x))

Merged Frangi, Koller images produce better TP/FN but increase FPs


Conclusion

• Lesser noisy connected pixel areas• Detects better the vessel width

Frangi• Detects better smaller vessels• Segments the bones thinner

Koller

Both delivers good results, better results maybe with optimized constants, post processing algorithms


Conclusion


Frangi Koller• Detects better smaller vessels• Segments the bones thinner


Conclusion




Conclusion



Documents

Extraction of Vessels from X-Ray Angiograms Titus Rosu Prof. Dr. Rupert Lasser Andreas Keil