Department of Ophthalmology, Perelman School of Medicine at the University of Pennsylvania

Transcription

1 Yuanjie Zheng 1, Dwight Stambolian 2, Joan O'Brien 2, James Gee 1 1 Penn Image Computing & Science Lab, Department of Radiology, 2 Department of Ophthalmology, Perelman School of Medicine at the University of Pennsylvania 1

2

3

4

5

6 Anterior Chamber Iris Lens Aqueous humour maintains a constant healthy eye pressure - continually produces a small amount of aqueous humor while an equal amount of this fluid flows out of the eye. flows through the narrow cleft between the front of the lens and the back of the iris escape through the pupil into the anterior chamber drain out of the eye via the trabecular meshwork (drainage canal).

7 When the drainage canal is blocked, fluid will build up and the pressure will increase.

8 Retina Increased pressure can damage the blood vessel and optic nerve on the retina.

9 Risk Factors Over the age of 40 Family history of glaucoma African-Americans Thinner corneas systemic vascular conditions (e.g. diabetes, hypertension, heart disease) prolonged corticosteriod use high myopia chronic ocular inflammation ocular trauma.

10 OCT Machine (Cirrus) OCT-Reconstructed Retinal Layers Provide cross-sectional image of retinal layers, which is analogous to a crosssectional histological section

11 ONH parameters automatically assessed using software tools from OCT merchant: cup area, disc area, rim area, cup-to-disc ratio thickness profile/map/symmetry of Retina Nerve Fiber Layer (RNFL)

12 Optical Nerve Head Iris Lens Retina Choroid Cornea creation of a photograph of the interior surface of the eye well-established gold standard CDR = Area cup /Area Disc CDR: cup-to-disc ratio

13 0.3CDR: normal

14 0.6CDR: moderate glaucoma 0.9CDR: severe glaucoma

15 Conventional Manual Depth analysis with stereo fundus pair Computerized planimetry Recent Automatic Template matching (Lalonde, et al., TMI, 2001) Machine learning (Abramoff, et al., IOVS, 2007; Xu, et al, MICCAI2012) Active contour model (Joshi, et al., TMI, 2011) Level set (Yu, et al., TITB, 2012) Hough transform (Aquino, et al., TMI, 2010) Challenges Blurry and faint boundaries Large inter-subject variability Interference of blood vessel Weakness Independently segment cup and disc Remove vessel with a preprocess Retinal Vasculature Temporal Side Optical Disc Center Optical Disc Boundary Optical Cup Boundary Nasal Side

16 Circular boundaries around disc center Cup and disc radii can be a priori estimated Cup in disc Rim thickness can be a priori estimated Cup Boundary Disc Boundary

17 Pictorial structures (Felzenszwalb & Huttenlocher, IJCV, 2005) Graph Cut (Delong & Boykov, ICCV, 2009) Graph Searching (Li, Wu, Chen & Sonka, TPAMI, 2006) Lagrangian Duality (Ulen, Strandmark and Kahl, TMI, 2013)

18 Graph cut combinatorial optimization + Priors Optimizability + Fidelity (Ulen, et al. TMI2013)

19 Data Term SmoothnessTerm ( p) ( ) Vp f : cost of assigning label f to pixel p Vpq fp, fq : penalty when neighboring pixels pq, have different labels. Boykov, et al. TPAMI, 2001

20 Data Term SmoothnessTerm Background disc cup Region Interaction Term

21 D X x x x c p c d p p d p p x p ( ) = log, : binary (cup) (disc) P: a Gaussian Mixture Model (vessel not pre-removed) Cup is contained in Disc Submodular (Boykov, TPAMI, 2001)

22 l { "cup","disc"} I, I : intensity of neighboring pixels p and q p q R R, : average Euclidean distance to optic center : priori estimated radius value pq l Submodular Boundaries snapped to higher image gradient Boundary circularity and the a priori estimated radii are enforced for cup & disc

23 W X X x x x c cd,, ( ) p c d pq p : p = p, q: binary d x q Impose it on all pairs of pixels between which spatial distance is below a priori known value (rim thickness) Submodular Thinnest rim thickness enforced (Ulen, et al., TMI, 2013) Region containment enforced again ( disc contains cup )

24 Data Term SmoothnessTerm Region Interaction Term Submodular terms A single graph cut (Delong & Boykov, ICCV2009) Optimal solution computed with an efficient manner

25 248 optical nerve head (ONH) photographs taken by a simultaneous stereo fundus camera (Zeiss 30 o ) 30 progressed glaucoma and 32 control subjects Size: 2392x2042 pixels; Resolution: 2.6x2.6μm 2 HD-OCT (Cirrus; Carl Zeiss Meditec. Inc.) with ONH parameters of disc area, rim area, cup area and cup-to-disc area ratio generated by the Cirrus HD-OCT software (version 5.0) Disc and cup margins were manually delineated from photographs of randomly chosen 15 glaucoma subjects and 16 control subjects

26 Comparison of ONH measurements generated by the merchant-provided software of the OCT machine and our algorithm regarding area (mm 2 ) of disc, rim and cup and C/D ratio.

27 Statistics of average distance between the subjectively assessed optic cup and disc margins and the automatically detected margins by our algorithm (OUR), the active contour model (ACM) and graph cut (GC). $*$ indicates the best performance value. Normal Glaucoma

28 A fairly general energy function that can combine shape priors and region/boundary interactions. Global optimization framework + Priors (reduce search space of feasible solutions) -> reliability + robustness Deformable segmentation (Zhang, et al., MedIA, 2012), Active Shape Model (Cootes, et al., CVIU, 1995), Superpixel Classifcation (Xu, et al, MICCAI2012) training is needed

29 A larger data set Comparison with more techniques Optic disc center detection

30 Yuanjie Zheng: James Gee: