Statistical models of the spine with applications in medical image processing

Transcription

1 Faculté Polytechnique Statistical models of the spine with applications in medical image processing ISP Seminars, UCL -April 17th, 2013 Fabian Lecron

2 A Statistical Model? Sub-space characterizing a variable X based on observations Relevant informationsabout the model Statistical limits of the subspace Mean of the sample How doesevolvex insidethe sub-space?

3 Advantages Conventional Radiography Relatively weak exposure to radiation Fast and not expensive widely used in ER Disadvantages Human body only in 2D (not always a disadvantage) Poor quality

4 Objectives of my thesis Statistical models can help to deal with poor quality images Application to the analysis of the spine on radiographs We llseehow to extract2d or 3D spineshapesto computeclinical indices

5 Statisticalmodel definedby PCA

6 Statisticalmodel definedby PCA

7 PresentationOverview I. Models Multilevel statistical shape model Machine learning-based model II. Vertebral Mobility Context Automatic vertebra detection Vertebra segmentation III. Scoliosis Context 3D reconstruction with a statistical shape model 3D reconstruction with a machine learning-based model IV. Conclusion

8 MODELS MULTILEVEL STATISTICAL SHAPE MODEL MACHINE LEARNING-BASED SHAPE MODEL What if the shape has a hierarchical structure? In biomedical applications, data can have a hierarchical structure Usual models do not represent the link existing between the items of a structure Example: evaluating school performance of students

9 MODELS MULTILEVEL STATISTICAL SHAPE MODEL MACHINE LEARNING-BASED SHAPE MODEL Multilevel Model, Global representationof the spine

10 MODELS MULTILEVEL STATISTICAL SHAPE MODEL MACHINE LEARNING-BASED SHAPE MODEL Thesemodelsare definedby statisticalhypothesis Statistical limits

11 MODELS MULTILEVEL STATISTICAL SHAPE MODEL MACHINE LEARNING-BASED SHAPE MODEL Shape modeling based on data separation A shape is modelizedgiven an hyperplanerepresenting a class of data : Feature space Machine learning method: One-Class Support Vector Machine (OCSVM)

12 MODELS MULTILEVEL STATISTICAL SHAPE MODEL MACHINE LEARNING-BASED SHAPE MODEL Separationdefinedwitha kernel The hyperplane is computed from a kernel function Only the inner productk x,x Φ,Φ is required k x,x : easyto definebut not easyto choose!

13 PresentationOverview I. Models Multilevel statistical shape model Machine learning-based model II. Vertebral Mobility Context Automatic vertebra detection Vertebra segmentation III. Scoliosis Context 3D reconstruction with a statistical shape model 3D reconstruction with a machine learning-based model IV. Conclusion

14 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION Vertebral Mobility Measuring the vertebra orientations in different positions

15 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION Mobility Analysis: a Fully Why is the mobility important? Automatic Approach Helpful for diagnosis of vertebral pains (ex: trauma) Some pathologies imply a decrease of mobility need to quantify Why a fully automatic approach? Quickly providing quantitative data No inter-operator variability Processing lots of images (ex: for medical research)

16 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION Measuring Protocol

17 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION Framework in Three Steps Vertebra Detection Vertebra Segmentation Mobility Analysis

18 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION CannyEdgeDetector

19 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION GeometricalDefinitionof a Corner

20 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION Support Vector Machine

21 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION Interest Point Description Descriptor: features about the module and the orientation of the gradient

22 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION Descriptors: SIFT and SURF SIFT Descriptor SURF Descriptor 128 features 64 features Invariant to scale, rotation and illumination Fast Invariant to scale and rotation Very fast

23 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION Experiments 49 radiographs: cervical vertebraec3 to C7 leave-one-out cross-validation Corner and vertebra detection rates Precision of classification Results depend on a parameter: window descriptor size

24 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION Experiments

25 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION Experiments

26 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION Active Shape Model Statistical Shape Model Model of greylevelvariation aroundthe shape

27 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION Initializationclose to the object Mean shape is placed near the detected vertebrae

28 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION The Model isdeformable The texture in the neighborhood of the landmarks is analyzed

29 VERTEBRAL MOBILITY CONTEXT AUTOMATIC VERTEBRA DETECTION VERTEBRA SEGMENTATION Vertebral Mobility: Conclusion Inter-operator variability on radiographs: 3,14 Our approach vs. expert landmarking: between 3,15 and 3,36 Our approach is as precise as a group of radiologists Limitation: all the processrelies on the good detectionof the edges

30 PresentationOverview I. Models Multilevel statistical shape model Machine learning-based model II. Vertebral Mobility Context Automatic vertebra detection Vertebra segmentation III. Scoliosis Context 3D reconstruction with a statistical shape model 3D reconstruction with a machine learning-based model IV. Conclusion

31 SCOLIOSIS CONTEXT RECONSTRUCTION - STATISTICAL SHAPE MODELS RECONSTRUCTION - MACHINE LEARNING-BASED SHAPE MODEL 3D Deformation Represented with Cobb angle: widely used a 2D Clinical Measure The interest of 3D clinical indices has been shown in the literature

32 SCOLIOSIS CONTEXT RECONSTRUCTION - STATISTICAL SHAPE MODELS RECONSTRUCTION - MACHINE LEARNING-BASED SHAPE MODEL Reconstructing the Spine from Why using radiographs? Radiographs Fast and not expensive CT-Scan: important exposure, lying position MRI: expensive, lying position Disadvantage: weak image quality, poor contrast What for? Scoliosis diagnosis Personalized Treatments Surgery planning

33 SCOLIOSIS CONTEXT RECONSTRUCTION- STATISTICAL SHAPE MODELS RECONSTRUCTION - MACHINE LEARNING-BASED SHAPE MODEL Optimization of two measures: Reprojection error General Principle Similarity of the deformed shape with the statistical model

34 SCOLIOSIS CONTEXT RECONSTRUCTION- STATISTICAL SHAPE MODELS RECONSTRUCTION - MACHINE LEARNING-BASED SHAPE MODEL 3D Representation with Anatomical Landmarks

35 SCOLIOSIS CONTEXT RECONSTRUCTION- STATISTICAL SHAPE MODELS RECONSTRUCTION - MACHINE LEARNING-BASED SHAPE MODEL Statistical Shape Model

36 SCOLIOSIS CONTEXT RECONSTRUCTION- STATISTICAL SHAPE MODELS RECONSTRUCTION - MACHINE LEARNING-BASED SHAPE MODEL Multilevel Model,

37 SCOLIOSIS CONTEXT RECONSTRUCTION- STATISTICAL SHAPE MODELS RECONSTRUCTION - MACHINE LEARNING-BASED SHAPE MODEL Convex Optimization SOCP (Second Order Cone Programming) formulation: Advantage: the solution is found very quickly

38 SCOLIOSIS CONTEXT RECONSTRUCTION- STATISTICAL SHAPE MODELS RECONSTRUCTION - MACHINE LEARNING-BASED SHAPE MODEL Very Fast Solving StatisticalShape Model Multilevel Statistical Shape Model,

39 SCOLIOSIS CONTEXT RECONSTRUCTION- STATISTICAL SHAPE MODELS RECONSTRUCTION - MACHINE LEARNING-BASED SHAPE MODEL Experiments 20 severecases: Cobb angle between44 and 70 (Sainte-Justine Hospital, Montréal) 25 cases withsurgicalinstrumentation presenceof discontinuitiesin the spine 17 vertebraeare reconstructed: T1 to L5 Reconstruction error: euclidean distance to a reference 3D model Reconstruction time

40 SCOLIOSIS CONTEXT RECONSTRUCTION- STATISTICAL SHAPE MODELS RECONSTRUCTION - MACHINE LEARNING-BASED SHAPE MODEL Reconstruction Error Experiments Plates Pedicles

41 SCOLIOSIS CONTEXT RECONSTRUCTION- STATISTICAL SHAPE MODELS RECONSTRUCTION - MACHINE LEARNING-BASED SHAPE MODEL Time of reconstruction Experiments

42 SCOLIOSIS CONTEXT RECONSTRUCTION- STATISTICAL SHAPE MODELS RECONSTRUCTION - MACHINE LEARNING-BASED SHAPE MODEL Experiments

43 SCOLIOSIS CONTEXT RECONSTRUCTION - STATISTICAL SHAPE MODELS RECONSTRUCTION- MACHINE LEARNING-BASED SHAPE MODEL One-Class SVM for 3D Reconstruction Reconstruction algorithm based on the minimization of two measures:,,,, 1

44 SCOLIOSIS CONTEXT RECONSTRUCTION - STATISTICAL SHAPE MODELS RECONSTRUCTION- MACHINE LEARNING-BASED SHAPE MODEL Two Particular Kernels Hyperplane is defined by a kernel A kernel is actually a similarity measure Proposed kernel: RBF :, Mahalanobis :,

45 SCOLIOSIS CONTEXT RECONSTRUCTION - STATISTICAL SHAPE MODELS RECONSTRUCTION- MACHINE LEARNING-BASED SHAPE MODEL Two Representations Representation with landmarks:,,,,, ) Articulated representation:,,,,,,,,,, )

46 SCOLIOSIS CONTEXT RECONSTRUCTION - STATISTICAL SHAPE MODELS RECONSTRUCTION- MACHINE LEARNING-BASED SHAPE MODEL Reconstruction Error Experiments Landmarks Articulated Representation

47 SCOLIOSIS CONTEXT RECONSTRUCTION - STATISTICAL SHAPE MODELS RECONSTRUCTION- MACHINE LEARNING-BASED SHAPE MODEL Experiments Comparison with Statistical Shape Model

48 SCOLIOSIS CONTEXT RECONSTRUCTION - STATISTICAL SHAPE MODELS RECONSTRUCTION- MACHINE LEARNING-BASED SHAPE MODEL Experiments Outliers in the sample Sensitivity

49 PresentationOverview I. Models Multilevel statistical shape model Machine learning-based model II. Vertebral Mobility Context Automatic vertebra detection Vertebra segmentation III. Scoliosis Context 3D reconstruction with a statistical shape model 3D reconstruction with a machine learning-based model IV. Conclusion

50 Conclusion Statistical Shape Model Multilevel Statistical Shape Model Machine Learning-based Shape Model Interest of these models to extract the shape of the spine in 2D and 3D These models allow to choose conventional radiography instead of more harmful or more expensive modalities

51 Vertebral Mobility Conclusion Contribution : Automatic analysis of cervical vertebral mobility The statistical shape model guides the segmentation Scoliosis in 3D Contribution : Interactive reconstruction based on statistical shape models Contribution : Robust reconstruction based on OCSVM Reduction of the human intervention

52 Models Future Work Can beappliedto other«objects» (e.g. organsof the humanbody, etc.) Can beusedas statisticaltoolsto studya pathology(e.g. evolutionof vertebral deformations over time) Vertebral Mobility Extension of the approach to other modalities(e.g. videofluoroscopic system)? Other descriptors? Scoliosis in 3D Whatabout EOS system? OCSVM: simpler similarities Kernel-PCA sothatthe OCSVM shapemodel canbeusedas a statisticaltool