ivana isgum
TRANSCRIPT
Automatic determination of cardiovascular risk from thoracic CT scans using a
coronary calcium atlas
Image Sciences InstituteUniversity Medical Center Utrecht
NFBI, 10. 09. 2009.
I. Išgum, M. Prokop, P.C. Jacobs, M.J. Gondrie,W.P.Th.M. Mali, M.A. Viergever, B. van Ginneken
Cardiovascular disease
Class of diseases that involve the heart or blood vessels
Usually used to refer to diseases related to atherosclerosis– Condition in which an arterial wall
thickens
Atherosclerosis Arterial calcifications
– High density lesions inside arteries visible in CT scans
Calcium score – Value representing amount of
arterial calcifications – Independent marker of
cardiovascular disease
Calcium scoring in clinical practice
Possible calcifications extracted by thresholding
Calcifications are – Identified manually– Subsequently automatically
quantified
Goal
Automatic detection and
quantification of coronary calcifications in thoracic CT scans from lung cancer screening to determine cardiovascular risk status
Cardiovascular disease in lung cancer screening
More heavy smokers affected by cardiovascular disease than by lung cancer
Dutch-Belgian lung cancer screening trial (NELSON):– ~1% subjects - detected lung cancer
(baseline) – ~3% subjects (0.5% fatal) - cardiovascular
event (10 months follow-up)
Challenges for automatic coronary calcium scoring Thoracic CT scans from screening
– Non-ECG synchronization Cardiac motion
ECG-synchronized
Not ECG-synchronized
Challenges for automatic coronary calcium scoring Thoracic CT scans from screening
– Non-ECG synchronization Cardiac motion
– Non-contrast enhancement Coronaries not visible
Non contrast enhanced Contrast enhanced
Challenges for automatic coronary calcium scoring Thoracic CT scans from screening
– Non-ECG synchronization Cardiac motion
– Non-contrast enhancement Coronaries not visible
– Low-dose Noise with same intensity as calcium
Pattern recognition system Candidate extraction
– Thresholding (130 HU)– 3D connected-component labeling– Discarding too small (noise) and too large
(bony structures) candidates Feature computation
– Position (6 features)– Texture and size (41 features)
Classification– Two stage classification (kNN with feature
selection)
Position features For accurate calcium detection
– Location of the coronaries very important
Coronaries not visible– Segmentation not feasible
Alternative – Segmentation other anatomical
structures to estimate location of the coronaries*
Time consuming, less accurate* Isgum et al; Med Phys 2007; 34:1450-1461
Position Features Calcifications in the coronaries
are visible Design coronary calcium atlas
– Estimate location of the coronaries using calcifications in them
– Estimate probability for spatial appearance of calcifications
Data 121 thoracic CT scans
– 16 x 0.75 mm collimation – 0.7 mm section thickness– 1 mm increment
121 scans
51for creating
coronary calcium atlas
70 for designing
pattern recognition system
35 training 35 testing1 atlas 50 references
elastically registered* (coronary calcium atlas)
elastically registered to compute position features* http://elastix.isi.uu.nl/
Coronary calcium atlas
Probabilistic segmentation Atlas with segmentation
Results on 35 test scans Background Calcium Total
Background 211,673 32 211,705
Calcium 35 90 125
Total 211,708 122
Automatic systemRe
fere
nce
stan
dard
72% sensitivity;0.9 false positives/scan
Risk categories based on the coronary score
5 standard risk categories based on the Agatston score in the coronary arteries*
Risk category
I(very low)
II(low)
III(moderate
)
IV(moderately
high)V
(high)
Agatston score 0 1-10 10-100 100-400 >400
* Rumberger et al; AJR 2003; 181(3):743-74
I II III IV V
I4 0 0 0 0
II3 5 0 0 0
III1 2 7 0 0
IV0 0 0 7 0
V0 0 0 0 6
Reference standardAu
tom
atic
resu
lt
83% - agreement
I II III IV V
I4 0 0 0 0
II3 5 0 0 0
III1 2 7 0 0
IV0 0 0 7 0
V0 0 0 0 6
Reference standardAu
tom
atic
resu
lt
83% - agreement 14% - one category off
I II III IV V
I4 0 0 0 0
II3 5 0 0 0
III1 2 7 0 0
IV0 0 0 7 0
V0 0 0 0 6
Reference standardAu
tom
atic
resu
lt
83% - agreement 3% - two categories off
14% - one category off
I II III IV V
I4 0 0 0 0
II3 5 0 0 0
III1 2 7 0 0
IV0 0 0 7 0
V0 0 0 0 6
Reference standardAu
tom
atic
resu
lt
Conclusion Automatic calcium scoring from
low-dose, non-ECG synchronized thoracic CT scans appears feasible
In lung cancer screening cardiovascular risk can be estimated automatically