Perspectives on Automatic Image Segmentation for Radiotherapy. Greg Sharp Oct 25, 2013

Transcription

1 Perspectives on Automatic Image Segmentation for Radiotherapy Greg Sharp Oct 25, 2013 Karl Fritscher, UMIT Vladamir Pekar, Philips Marta Peroni, PSI Nadya Shusharina, MGH Harini Veeraraghavan, MSKCC Jinzhong Yang, MDACC

2 Outline Overview of segmentation algorithms Evaluation metrics Performance limits Outlook

3 Commercial products Company Product Method TPS integration Accuray MultiPlan Autosegmentation Atlas-based / model-based BrainLab iplan Atlas-based Yes Dosisoft IMAgo Atlas-based Yes Elekta ABAS Atlas-based / model-based MIM MIM Maestro Atlas-based No Mirada RTx Atlas-based No Philips SPICE Atlas-based / model-based RaySearch RayStation Atlas-based / model-based Varian Smart Segmentation Atlas-based Yes Velocity VelocityAI Atlas-based No Yes No Yes Yes ** WIP. Please notify Greg of corrections.

4 Atlas-based segmentation Atlas DIR Subject Warped contours

5 Multi-atlas segmentation... Atlas 1 Atlas 2 Atlas n DIR DIR DIR Subject Warped contours Label Final fusion segmentation Warped contours Warped contours

6 Atlas selection ** Rohlfing, 2005

7 Subject... Pre- selection DIR DIR

8 Subject... Pre- selection DIR DIR Post- selection

9 Atlas selection Image similarity Mutual information, SSD, etc. Registration Rotation, skew, deformation, etc.

10 Label fusion Voting Majority voting, weighted voting, etc. STAPLE Spatial regularization, MAP-STAPLE, etc.

11 STAPLE Segmentations by different raters p = 0.5 q = 0.5 p = 0.7 q = 0.9 p = 0.8 q = 0.6 Raters' sensitivity & specificity True segmentation

12 STAPLE Segmentations by different raters p = 0.5 q = 0.5 p = 0.7 q = 0.9 p = 0.8 q = 0.6 Raters' sensitivity & specificity E-Step For each voxel, estimate likelihood the voxel belongs to true segmentation

13 STAPLE Segmentations by different raters Estimate of true segmentation M-Step p = 0.5 q = 0.5 p = 0.7 q = 0.9 p = 0.8 q = 0.6 Re-estimate rater sensitivity & specificity

14 Limitations of atlases Imprecision of registration "Averaging" effect Sometimes visually unappealing

15 Model-based segmentation Active shape model ** Cootes, 1995

16 Model-based segmentation Active shape model P Position of point P varies with a known statistical distribution over a population These statistics are used to control the range of allowed shapes ** Cootes, 1995

17 Model-based segmentation Active appearance models Model of intensities Statistical deformation models Model of deformation vectors InShape models Joint model of intensity and deformation vectors

18 Model-based segmentation Brainstem ** Fritscher, 2013 Left Parotid Physician Atlas Atlas + InShape contour model

19 Dice's coefficient Invented in 1940's for botany applications Dice(A, B)= 2 A B A + B

20 Dice's coefficient Invented in 1940's for botany applications Dice(A, B)= 2 A B A + B

21 Dice's coefficient What is the Dice coefficient when matching a parotid gland with a sphere?

22 Dice's coefficient What is the Dice coefficient when matching a parotid gland with a sphere?

23 Dice's coefficient What is the Dice coefficient when matching a parotid gland with a sphere?

24 Dice's coefficient What is the Dice coefficient when matching a parotid gland with a sphere? Dice coefficient = /- 0.05

25 Hausdorff distance Image credit:

26 WARNING Professional Driver. Closed course. Photo credit:

27 Hausdorff distance One-sided Hausdorff distance Hausdorff 1 (A, B)=max a A min a b b B Average Hausdorff (take one) Hausdorff Ave ( A, B) = 1 2 Hausdorff 1( A, B) Hausdorff 1(B, A)

28 Hausdorff distance One-sided Average distance Hausdorff 1, Ave ( A, B)= 1 A a A Average Hausdorff (take two) min a b b B Hausdorff Ave ( A, B)=max( Hausdorff 1, Ave( A, B) Hausdorff 1, Ave ( B, A)) Or is it the average of the two instead of max??

29 Hausdorff distance One-sided fractional (95%) Hausdorff Hausdorff 1,95 (A, B)=P 95 a A( min b B a b )

30 Hausdorff distance One-sided fractional (95%) Hausdorff Hausdorff 1,95 (A, B)=P 95 a A( min b B a b ) Fractional (95%) Hausdorff Hausdorff 95 (A, B)=max( Hausdorff 1,95 ( A, B) Hausdorff 1,95 (B, A))

31 Hausdorff distance One-sided fractional (95%) Hausdorff Hausdorff 1,95 (A, B)=P 95 a A( min b B a b ) Fractional (95%) Hausdorff Hausdorff 95 (A, B)=max( Hausdorff 1,95 ( A, B) Hausdorff 1,95 (B, A)) Or is it the average? Or should I combine the points, then take 95%?

32 Here be dragons Photo credit: Public Domain

33 Boundary Hausdorff Hausdorff distance may be computed on the set or the set boundary

34 Boundary Hausdorff The max distance is to a point on the boundary So no difference, right?

35 Boundary Hausdorff The max distance is to a point on the boundary So no difference, right? Average distance will change Hausdorff Ave ( A, B)= 95% distance will change Hausdorff 95 ( A, B)=

36 Boundary Hausdorff "max" Hausdorff changes too Hausdorff ( A, B) Hausdorff ( A, B)

37 Boundary Hausdorff "max" Hausdorff changes too Hausdorff ( A, B) Hausdorff ( A, B) Hausdorff ( A, B) Hausdorff ( A, B)

38 Don't try this at home! Photo credit:

39 Average Dice How should I get a single score for a segmentation of many structures? Should I average Dice over all structures? S structures Dice (S, Ŝ )

40 Average Dice

41 Dice vs Hausdorff Dice and Hausdorff measure different things. Should I average them too?

42 Dice vs Hausdorff Dice and Hausdorff measure different things. Should I average them too?

43 Inter-observer variability Structure Inter-observer variability Parotid gland 0.66 ± 0.1 [1] 0.76 ± 0.08 [2] 0.85 [3] Automatic segmentation accuracy 0.74 [4] [0.73,0.79] [5] 0.85 ± 0.03 [6] ** All scores are Dice [1] Sims, 2009 [2] Nelms, 2012 [3] Faggiano, 2011 [4] Mattiuchi, 2013 [5] La Maccia, 2012 [6] Pekar, 2010

44 RTOG 0522

45 RTOG 0522

46 Inter-observer variability GCS / KDF / GCS / RTOG / 0.09 KDF / RTOG / 0.12 ** GCS & KDF used protocol described in van de Water et al., 2009

47

48

49

50

51 Outlook Short term RTOG delineation guidelines become common Open databases / open-source implementations Increased use of shape models Medium term Move from atlas-based to machine learning-based Increased need for multi-modal segmentation Segmentation of treatment target volumes