Automatic Detection and Segmentation of Ground Glass Opacity GGO Nodules

1
Automatic Detection and Segmentation of Ground
Glass Opacity (GGO) Nodules

• Jinghao Zhou, Sukmoon Chang, Dimitris Metaxas
• Center for Computational Biomedicine Imaging and
  Modeling (CBIM)
• Rutgers, The state University of New Jersey

2
Background

• Ground Glass Opacity (GGO)
• Hazy increased attenuation within a lung
• More likely to be malignant than solid opacity
• Represent active disease
• Pulmonary edema
• Pneumonia
• Diffuse alveolar damage
• Early sign of bronchioloalveolar carcinoma (BAC)
• Detection and treatment of pure GGO can improve a
  prognosis of lung cancer

3
Existing methods

• Hybrid neural network
• Three single nets and an expert rule for GGO
  detection
• Underestimates GGO area
• Due to its improper cut-off of the edges of GGO
• May be used only for large GGO
• Inaccurate segmentation for small GGO

4
Existing method

• Automatic clustering techniques
• GGO detection only
• Markov random field
• Vessel removal method based on shape analysis
• GGO segmentation only
• Manual identification of GGO

5
Difficulty

• Appearances of GGO
• Irregular shape
• Fuzzy boundary
• Overlapping vessels

6
Proposed method

• Two-step process
• GGO detection
• Based on a machine learning framework
• GGO segmentation
• Base on a nonparametric density estimation

7
GGO Detection

• Detect candidate GGO areas
• Cylinder filter for vessel/noise suppression
• Thresholding
• Classify candidate GGO areas
• k-NN on intensity probability density functions
  (p.d.f.)
• Boosting k-NN for speed-up

8
Candidate GGO Area Detection

• Hybrid neighborhood cylinder filter
• Strong responses to blob-like objects (GGO)
• Suppresses tubular objects and noise

O Domain of the cylinder x, ? Center and
orientation of the cylinder
9
Candidate GGO Area Detection

• Vessel / noise suppression
• Apply cylinder filters of 7 orientations
• Cylinder size used
• Radius of 1, 2, 3 voxels
• Length of 7 voxels

10
Candidate GGO Area Detection

• Result of the cylinder filters

11
Candidate GGO Area Detection

• Isolate candidate GGO areas
• Simple thresholding
• Histogram analysis of the filtered 3D dataset
• Automatic threshold selection

Histogram of the 3D cylinder filtered dataset.
12
Candidate GGO Area Detection

• Result of thresholding

One slice of original 3D dataset
Same slice after cylinder filtering
Same slice after thresholding
The vessels and noise are effectively suppressed
while GGO remains.
13
GGO Classification

• Collect example images
• Positive and negative examples

Negative examples (Non GGO)
Positive examples (GGO)
14
GGO Classification

• Estimate image intensity p.d.f.
• Nonparametric density estimation

i Random variable for intensity values.
i0,,255. AM 2D image examples bounded by a
square model ?M S(AM) Area of AM ,
y Pixels in the domain AM s Standard
deviation of Gaussian Kernel.
15
GGO Classification

• Typical p.d.f. of example images

16
GGO Classification

• Classification of GGO with k-NN
• Find k closest examples to an input image
• Classify the input image
• As the class to which the majority of the k
  closest examples belong to
• Disadvantage
• Requires a large training set of samples
• Classification is slow

17
GGO Classification

• Classification of GGO by boosting k-NN
• Run k-NN repeatedly
• Select the best examples for classification
• Combine the examples selected
• Advantage
• Reduce the number of examples for classification
• Speed up the classification process

18
GGO Classification

• Classification Results
• 319 example images
• 200 training / 119 testing examples
• Each example 99 pixels
• Generalization error estimation
• Using Bootstrapping examples
• 20 steps of boosting for classification
• Mean error rate 5.04

19
GGO Classification

• Comparison to other learning methods

20
Proposed method

• Two-step process
• GGO detection
• Based on a machine learning framework
• GGO segmentation
• Base on a nonparametric density estimation

21
GGO Segmentation

• Extract ROI (region of interest)
• 3D ROI around GGO
• Estimate nonparametric density
• On 3D GGO volume
• On neighboring voxels in ROI
• Compute 3D likelihood map
• Likelihood of each voxel belonging to GGO
• Remove overlapping vessels

22
GGO Segmentation

• Extract ROI
• Around the center of mass of GGO

RM 3D GGO volume bounded by a cubic
model FM
RLoc 3D sphere volume of a neighboring
voxel in ROI bounded by a model FLoc
23
GGO Segmentation

• Estimate nonparametric density
• On 3D GGO volume

i Random variable for intensity values.
i0,,255 RM 3D GGO volume bounded by a cubic
model FM V(RM) Volume of RM, y Pixels in
the domain RM s Standard deviation of Gaussian
Kernel
24
GGO Segmentation

• Estimate nonparametric density
• On 3D sphere volume of each voxel in ROI

i Random variable for intensity values.
i0,,255. RLoc 3D sphere volume of a voxel in
ROI bounded by a model FLoc V(RLoc) Volume of
RLoc, y Pixels in the domain RLoc s
Standard deviation of Gaussian Kernel
25
GGO Segmentation

• Compute 3D likelihood map
• Likelihood map of a voxel belonging to GGO

p1 P(i FM ), p2 P(i FLoc ) B(p1, p2)
Bhattacharya distance between the two
p.d.f.s ?(p1, p2) 3D likelihood of a voxel in
the 3D ROI belonging to the GGO
26
GGO Segmentation

• Compute 3D likelihood map
• Likelihood of a voxel belonging to GGO

One slice of original volume data of GGO
Same slice of 3D likelihood map of GGO
27
GGO Segmentation

• Remove overlapping vessels
• Eigenanalysis of the Hessian Matrix

28
GGO Segmentation

• Remove overlapping vessels

One slice of 3D likelihood map of GGO
Same slice after vessel removal
29
GGO Segmentation

• Results

30
GGO Segmentation

• Results

31
GGO Segmentation

• Results

32
Conclusions

• Difficulties in GGO processing
• Irregular shape, Fuzzy boundary
• Overlapping vessels
• Proposed method
• Two-step process
• GGO detection by machine learning framework
• GGO segmentation by nonparametric density
  estimation
• Automatic GGO processing
• Accurate detection
• Reproducible segmentation

33
Thank you...

• Questions?