First make a centerline representing the artery. Simpler to make measurements on. Find end-points to measure from.
Slab ends at variable point. Tortuosity measurement can be taken at peak or end of curves.
Higher peaks for more tightly wound coils. Oscillating shapes create oscillating curve.
Radio frequency coils generate signal. Gradient coils encode spatial position.
Segmentation separates flowing arterial blood from stationary background tissues.
Cast rays through 3D data and display position of brightest point on each ray. Arterial blood is smooth in the image. MIP-Z smoothness defines a set of seed points; not full 3D artery segmentation.
Slow moving or recirculating blood in aneurysms have low signal; appear as background.
Hole filling especially needed in aneurysms. Aneurysm is a dilation 1.5 X vessel diameter. Holes touching outside aren’t filled in by connected component bubble filling.
Compare centerline algorithms used for anatomy assessment.
How we make a centerline. Cost function applied to segmentation has to be cheap in middle and expensive outside. Least cost centerline goes to middle. Working from the goal node assign the least cost back to the goal node from every voxel in the segmentation. Next slide describes removing short paths.
Optional cost function. MDFE higher in middle; lower on outside. Needs reversing.
Centerline will go to low cost middle.
Black area in middle actually has a gradient of values.
Dim short branches were pruned by shortest paths centerline algorithm.
Compare algorithm stability starting from different goal nodes. Phantom generated starting with lines of dots and fill in around dots. Original dots used as true centerline.
Green known centerline. Red is calculated centerline missing green. Yellow is overlap between known and calculated. Brighter stability plot; all centerlines not taking the same path. Display scales stability intensity.
BT-DFE and BT-COM are BT eroded data input into other algorithm. The stability measure for an image was the percentage of centerline voxels in the accumulated image called centerline for all of the centerline roots. Stability is fraction of all points that are the same from all starting points.
Only COM doesn’t have errors in ICA siphon loop.
Sometime the MDFE is correct but not from all goal nodes.
BT eroded data so few alternatives exist. BT is inherently stable.
Apply centerline hypertensive population
Made phantom to challenge COM algorithm. Weighted COM with DFE to make voxels toward middle have more weight in centerline calculation. COM centerline pulled to one side.
Humans are more similar to each other than to computer. Repeated experiment and got lower correlations between neurosurgeons.
Hypertensives have less microvessels.
Images not all at same resolution. Double resolution increases tortuosity about 5%. Closer resolutions more similar tortuosity scores. 0.23x0.23x0.36
DFM curve was good enough to show statistical significant difference, but not clinically useful due to overlap. Hypertension can be used as a training set testing tortuosity measurements to increase separation between groups to find clinically significant measure. Phase frequency artifact. Pulsatile flow. X and Y position are recorded at different times.
Repeat experiment with Utah population. Utah and North Carolina negatives similar. Shows that Utah hospital control of patients with headaches or head injuries are a valid negative control. Difference not explained by sex or age. Ethnicity is different. Utah and NC are both mostly white European populations. Use specific negative controls for each test population.
Only compared within Utah population. Utah hypertensive population on hypertensive medication.
Highest, median and low tortuosity subjects all have intracranial aneurysms. Marfan syndome can be misdiagnosis of Loeys-Dietz syndrome.
Compared Aneurysms, high-risk aneurysms, high-risk no aneurysms versus Utah negative control.
Database and plotting interface allow distribution viewing. Arnold-Chiari malformation: structural defects in the cerebellum, the part of the brain that controls balance Combination of tortuosity and medical record screening for Marfan, Arnold-Chiari malformation can identify LDS plotDFM(pwd=kpwd, conType='RODBC', arteryIds=c(5), cmdline=TRUE, legendx=.5, legendy=.95, hist=TRUE)
Biomedical informaticians always have to talk about what biomedical informatics is.
Aterial tortuosity mesurement system
ARTERIAL TORTUOSITY MEASUREMENT SYSTEM FOREXAMINING CORRELATIONS WITH VASCULAR DISEASE Karl Diedrich
Compare vascular disease to negatives No vascular disease Vascular Disease High risk aneurysm relative (10% risk) Normal aneurysm risk (5%) Aneurysmsuruda, D. Parker, J. MacDonald, and L.A. Cannon-Albright, “Confirmation of chromosome 7q11 locus for predisposition to intracranial aneurysm,” Hu 2
Centerlines with bifurcation guidesenterline bifurcations guide Anterior Cerebral artery (ACA) centerline selec selection of end points Cross section Projection 3
Tortuosity measuremente Factor Metric (DFM) = Length(L)/distance between end MCA-ACA bifurcation L d Internal carotid artery End of slab Repeated measurements, same patient 4
Medical image segmentation Z-Buffer segmentation  of arteriestic Resonance Angiography images highlight flowing arterial bloodL. Alexander, and J. S. Tsuruda, “Enhanced image detail using continuity in the MIP Z-buffer: applications to magnetic resonance angiography,” Journal of Magnetic Resonance Ima 8
MIP Z-buffer segmentation Intensity is position in image slice stack of maximum pixel intensity; dark is closer, brighter is farther Contiguous blood vessels are smoothd J. S. Tsuruda, “Enhanced image detail using continuity in the MIP Z-buffer: applications to magnetic resonance angiography,” Journal of Magnetic Re 9
2-D seed imageOriginal intensityvalues for smoothclusters over thethresholdUsed as seeds togrow 3-D imagefrom 10
Seed histogram thresholdHistogram of 2-D seed20% of histogram from the left is used to findintensity threshold for 3-D region growing Count 20% below 135 Intensity value 11
3-D Region GrowingCheck if pixelsneighboring 26 voxelsare above seedhistogram thresholdand add non-maximal3-D pixels 12
Region growing thresholdLowering region growing in 26 directions threshold0.20 histogram seed threshold 0.07 histogram seed threshold Noise Aneurysm 0.20 histogram threshold slice 0.07 histogram threshold sli 3T 13
Hole Fill No filling Bubble fillingBubble filling usesconnected components tofill bubbles completelyenclosed bubbles inaneurysmVoxel filing fills in individualvoxels with arteryneighbors in (variable) 24of 26 directions within 8 Voxel filling Bubble + voxel fillingvoxelsBubble fill -> 3 voxel fills ->bubble fill 1.5 T scanner, region growing >= 0.20 14
Paper 1Centerline Algorithms for Quantitative AssessmenKarl T. Diedrich, John A. Roberts, Richard H. Schmidt and Dennis L. Parker 15
Least cost path centerline Cost functionsGoal node Cross section Least cost paths back to goal node voxel Backtrace from distal ends to goal and remove s 16
Centerline Path costsL. Zhang et al., “Automatic detection of three-dimensional vascular tree centerlines and bifurcations in high-resolution magnetic resonance Goal node Removed short path This path made first Branch meets previous line 17
Distance From Edge (DFE)Pythagorean theorem d2 = x2 + y2 + z2 d y x Diagonal distances are longer than straight 18
M Modified Distance From Edge (MDFE)Increase MDFE of central voxels (V).MDFE(Vi) = DFE(Vi) + N(Vi)/NmaxN(Vi) = neighbor voxels with same DFENmax = possible neighbours CrossCenter voxel has same DFE in Z sections DF MD Higher intensity in image is higher value FE 19
Inverse cost functionCost(Vi) = A * (1 - MDFE(Vi)/max_MDFE(Vi) )b +1Inverts to make lower cost internal Lower intensity lower cost Inversion cost function MD Cost 20
Modified Distance From Edge (MDFE) MDFE cross section 21
Center of mass movement Segmentation Mean x, y, z position of each voxel, Vi, and up to 26 neighbors; R Segmentation collapsing to center of massAccumulate the distance moved 22
Center of mass costcost is the total distance move. Exterior voxels move farther to COM; highe 23
Binary thinned arterydes segmentation to single lines. Pass to centerline algorithm toht Journal - Implementation of a 3D thinning algorithm,” 12-Oct-2007. [Online]. Available: http://www.insight-journal.org/browse/publication/181. [A 24
Multiple centerlines stability test Second round goal nod COM First goal node 25
Phantom stability & accuracy A-B) MDFE C-D) COMability, brighter centerline Green known centerline. E-F) BT-MDFE G-H) BT-COM Red calculated centerline. Yellow is overlap. Stability Accuracy 26
Helix and line phantomRoot Mean Square Error (RMSE) of accuracy. Lower is better.A lg orith m S ta b ility R MS E of Ac c u ra c yMD F E 0 .8 8 0 0 .2 4 0C OM 0 .9 8 0 0 .6 1 0B T-MD F E 1 .0 0 0 1 .8 3 3B T-C O M 1 .0 0 0 1 .8 3 0 27
Artery centerline stability A) MDFE B) MDFE C) COM D) COM E) BT-COM F) BT-COM Arrows show errors in ICA siphon loop 28
Artery centerline stabilitystability compares well with inherently stable BT algorithms (8 su 29
Kissing vessels (ICA)DFE cost cross section Kiss COM cost cross section Kiss Segmentation Kiss MDFE costM cost, completes loop Binary thinned 30
Stability of arterial centerlinesA lg orit IC A Portion B oth Me a n S ta n d a rd Me a n S ta n d a rdhm s ip h on s IC A IC A num ber d e v ia tion s ta b ili d e v ia tion a c c u ra te s ip h on s c orre c t of tre e s of tre e s ty s ta b ility c orre c t in im a g eMD F E 6 /1 6 0 .3 7 5 1 /8 3 8 .8 7 5 1 4 .6 7 2 0 .6 7 7 0 .0 7 6C O M 1 6 /1 6 1 .0 0 0 8 /8 3 5 .1 2 5 1 3 .3 1 4 0 .8 7 7 0 .0 4 2B T- 1 0 /1 6 0 .6 2 5 4 /8 3 7 .5 0 0 1 3 .6 1 7 0 .8 8 3 0 .0 6 8C OM 31
Paper 2of an arterial tortuosity measure with application to hypertensio 32
Lopsided phantom accuracyd phantom challenges COM COM MDFE DFE-COM A lg orith m N u m b e r of tre e s S ta b ility R MS E of Ac c u ra c y C OM 6 0 .9 1 8 0 .8 7 9 MD F E 6 0 .8 1 9 0 .4 1 7 D F E -C O M 6 0 .9 0 5 0 .4 1 3 33
DFE-COM ICA siphonA lg IC A Portio B oth Porti Me a n S ta n d a r Me a n S ta n d aorit s ip h n IC A IC A on num d s ta b ili rdhm on s s ip h o c orre c orre ber d e v ia ti ty d e v ia ti accu ns c t in ct of on of on ra te c orre im a g im a g tre e s tre e s s ta b ilit ct e es yC O M 1 5 /1 6 0 .9 3 8 7 /8 0 .8 7 5 3 7 .0 0 1 2 .3 5 2 0 .8 7 2 0 .0 4 5 9 0MD F 7 /1 6 0 .4 3 8 1 /8 0 .1 2 5 3 9 .8 7 1 3 .2 2 8 0 .6 7 3 0 .0 7 3 2E 5D F E - 1 5 /1 6 0 .9 3 8 7 /8 0 .8 7 5 3 8 .6 2 1 1 .4 3 9 0 .8 2 5 0 .0 4 3 4C OM 5 34
Visual versus quantitative ranking DFM to mean human 0.72 Spearmen rank c Between humans 0.88±0.048 25 arteries 5 observers 35
Hypertension in microvessels HTN NOReries (LSA) in hypertensives (HTN) increased tortuosity, less number than normotensives (NOR) (7 Tg et al., “Hypertension correlates with lenticulostriate arteries visualized by 7T magnetic resonance angiography,” Hypertension, vol. 54, no. 5, pp. 10 36
Resolution effect on tortuositySame subjects at different resolutions by acquisition and interpolation 37
Hypertension and tortuosityA rte ry P-v a lu e L e ft AC A 0 .0 0 3 7 7 R ig h t AC A 0 .0 5 9 3L to R AC A 0 .0 1 6 5 L e ft IC A 0 .0 2 1 5 R ig h t IC A 0 .1 4 2 L e ft L S A s 0 .0 0 1 6 1R ig h t L S A s 0 .0 0 0 5 2 0 L e ft L S A s 0 .0 0 9 7 7R ig h t L S A s 0 .0 0 0 8 0 0 L e ft L S A 1 0 .0 2 3 8R ig h t L S A 1 0 .0 0 9 0 5 L e ft L S A 1 0 .0 8 8 0R ig h t L S A 1 0 .0 7 8 6 HTN N = 18±3.0 NEG N = 18±3.8 1-sided Wilcoxon signed rank test 38
Negative controls Korean negative control consistently lower Utah hospital same as North Carolina negative controlffects of healthy aging on intracerebral blood vessels visualized by magnetic resonance angiography,” Neurobiology 39
Utah hypertensionNone significant at α = 0.05Utah hypertensives on anti-hypertensive medication 40
Paper 3d D. L. Parker, “Medical record and imaging evaluation to 41
Aneurysms and tortuosity A rte ry P-v a lu e L e ft AC A 0 .0 0 0 5 4 R ig h t 0 .0 7 9 AC A L to R 0 .3 2 0 AC A B a s ila r 0 .1 5 7 L e ft IC A 0 .0 9 7 R ig h t 0 .0 7 8 IC A L e ft VA 0 .0 4 3 R ig h t VA 0 .4 3 1urysm N = 53±10ative N = 36±5.9ded Wilcoxon signed rank test 43
Loeys-Dietz tortuosity A rte ry P-v a lu e AC A le ft 0 .4 7 4 AC A 0 .1 3 1 rig h t B a s ila r 0 .0 0 4 5 0 L -R AC A 0 .0 6 3 1 IC A le ft 0 .3 2 2 IC A 0 .2 1 6 rig h t V A le f t 0 .0 0 0 4 3 VA rig h t 0 .0 5 0 9 Dietz N = 4.5±1.2 ve N = 36±5.9d Wilcoxon signed rank testtially distinguish LDS from Marfan with tortuosity 44
Tortuosity distribution Arnold-Chiari malformation: occurs 1 in 1280, 13.3% of LDS patients  Marfan diagnosis: LDS can be misdiagnosed as Marfan Loeys-Dietz (LDS) mean = 1.9 Collection of negative controls and vascularomes caused by mutations in the TGF-beta receptor,” The New England Journal of Medicine, vol. 355 45
Components of medical informatics Signal processing Applied image processing to anatomical measurement Database design 5/5 Applied database design to medical image analysis Decision making Aided diagnosing Loeys-Dietz syndrome Modeling and simulation Simulated artery shapes to challenge centerline algorithms Optimizing interfaces between human and machine Artery and centerline measurement and display Centerline visualizationsMedical informatics: a real discipline?,” Journal of the American Medical Informatics Association: JAMIA, vol. 2, no. 4, pp. 207-21 46
Experiment conclusionsMethods detected increased arterial tortuosityHypertensive sampleLoeys-Dietz syndrome sampleIncreased tortuosity could distinguish Loeys-Dietz from related MarfanCorrelated Loeys-Dietz syndrome TGFBR2genotype with tortuosity phenotype 47
System conclusionsFlexible analysis systemChange groups in comparisonsChange and modify tortuosity algorithmsReanalyze with new dataSecondary use of existing imagesEnabled by interpolation of imagesEnables quick less expensive testing of hypothesesUse to decide on best prospective studies 48
AcknowledgmentsCommittee: John Roberts, Richard Schmidt, Lisa Canon-Albright, Paul Clayton, Dennis ParkerCo-authors: John Roberts, Richard Schmidt, Lisa Canon-Albright, Dennis Parker, Chang-Ki Kang, Zang-Hee Cho,Anji T. YetmanThis work was support by NLM Grants: T15LM007124,and 1R01 HL48223, and the Ben B. and Iris M. MargolisFoundation.Many thanks to the students and staff at Utah Centerfor Advanced Imaging Research (UCAIR) 49
Acknowledgmentsnstitute (NRI), Gachon University of Medicine and Science , Division Of Cardiology, Primary Childrens Medical Cent, University of Utahn and Leo 50