miccai 2015
TRANSCRIPT
Joel Carlson, Sung-Joon YeMSc StudentRadiological Physics LabSeoul National University
A Radiomics Approach to the Classification of Astrocytomaand Oligodendroglioma
http://joelcarlson.me @jnkcarlson
ContentsMotivationMethods
Types of features calculatedImage processingTraining sample creationModel training and validation
Exploration of ResultsConclusions
Is it reproducible?Not really.
-Concordance index among 4 experienced neuropathologists: 52-70%
The Challenge and ProblemCorrect classification of glioma category essential for physicians and patients:
-Therapy choices-Understanding prognosis
Oligodendroglioma
Astrocytoma
*
*
Current Method:Neuropathologists determine classification
-Based on biopsy specimen-Relatively subjective
ContentsMotivationMethods
Types of features calculatedImage processingTraining sample creationModel training and validation
Exploration of ResultsConclusions
Types of FeaturesWndchrm
-Raw image features-Transformed image features-Compound transform features
Texture via Radiomics R package-First order features (energy, entropy, etc)-GLCM, GLRLM (rotation invariant)-Thibault matrices: GLSZM, MGLSZM
Nuclei features -Area, perimeter, max. diameter, skewness, kurtosis, etc)
Raw Fourier Transform
LoG
Edge
Pathology: Regions of Interest
Raw image
Superior method: Tiling
Select 5 ROIs
For each ROI calculate wndchrm features
Steps:1. Raw image (each pathology ROI)2. Color Deconvolution
a. Eosinb. Hematoxylin
3. Thresholding– Gaussian blur– Fill holes– Watershed
4. Nuclei detection– Area greater than some constant
5. Apply mask to raw image6. Calculate Nuclei features
Pathology: Nuclei Segmentation2a 2b1
3
4
5
Radiology: Extraction of tumor components
BraTumIa used to segment tumor1. Mask raw image with classification2. Extract masked region for each
component
3. Calculate Features– Wndchrm– Textures
For the slice that most heavily expresses a given component:
Model Building
3 Phases in a loop:
1. Create random samples:– Patients (training/validation)– Data
2. Permute data and combine; train models for each permutation
3. Validate models; make and save predictions
Phase 1: Sampling Patients and Data
Sample Training and validation patients:
Sample data types:
Radiological
Full Tumor
Edema
Necrotic
Enhancing
Non-Enhancing
T1c
FLAIR
Image Type:Choose 1
Tissue Type:Choose 1
Nuclei
Pathological
1
2
3
4
5
ROI:Choose 1
Training • 75% of patients
Validation • 25% of patients
Phase 2: Creating Permutations and Training
*Generalized Linear Model (LOOCV to determine regularization)
Rwnd_Rtext
Rwnd_Pwnd
Rwnd_Nshape
Rtext_Pwnd
Rtext_Nshape
Nuclei
Pwnd_NShape
RadiologicalTexture
Pathologicalwndchrm
Radiologicalwndchrm
NucleiShape
Train Model for each
permutation
*
Pre-Process(Center, Scale,
PCA)
Make Test Set Predictions
Make Validation Set Predictions
Phase 3: Validation and PredictionsFor each model trained:• Calculate Cross validation accuracy on validation set (25% of patients)• Save predictions of each model on testing set
Final Predictions• Majority class as voted by
models with:• CV accuracy > 85%• More than 3 variables
included in model
ContentsMotivationMethods
Types of features calculatedImage processingTraining sample creationModel training and validation
Exploration of ResultsConclusions
Results – CV Accuracy Density Histograms
Mean CV Accuracy: 0.502
Only Non-Enhancing shows predictive ability greater than chance on CV accuracy
Models with CV Acc > 0.85 used for predictions
1
Exploring the Non-Enhancing CV AccuracyMean CV Acc: 0.502
All Models
1.Mean CV Acc: 0.608NonEnhancing
Exploring the Non-Enhancing CV Accuracy
2Mean CV Acc: 0.502
All Models
1.Mean CV Acc: 0.608NonEnhancing
2.Mean CV Acc: 0.651NonEnhancing, Select Models
Exploring the Non-Enhancing CV Accuracy
3Mean CV Acc: 0.502
All Models
1.Mean CV Acc: 0.608NonEnhancing
2.Mean CV Acc: 0.651NonEnhancing, Select Models
3.Mean CV Acc: 0.688NonEnhancing,Select Models,
T1c only
4.Mean CV Acc: 0.734NonEnhancing,
Rwnd_Rtext only,T1c only
Exploring the Non-Enhancing CV AccuracyMean CV Acc: 0.502
All Models
1.Mean CV Acc: 0.608NonEnhancing
2.Mean CV Acc: 0.651NonEnhancing, Select Models
3.Mean CV Acc: 0.688NonEnhancing,Select Models,
T1c only
4
ConclusionsRadiological WNDCHRM and Texture features provide useful and orthogonal information
-Nuclei features show some promise (not discussed)
Majority of features not useful
Using PCA obfuscates which features may be useful
Better predictions may take into account dependence of CV accuracy on number of variables
Mean CV accuracy of all models with >3 variables: ~0.5 Mean CV accuracy of select models* with > 16 variables: ~0.7
*Certain T1c Non-Enhancing models (Rtext, Rwnd, Rwnd_Rtext)
http://joelcarlson.me @jnkcarlson
Thank you for your attention!
Four categories• High contrast features
• Edges, connected components, spatial distribution, size and shape
• Polynomial decompositions• A polynomial that approximates
the image to some fidelity is generated. Coefficients used as descriptors.
• High contrast features• Edges, connected components,
spatial distribution, size and shape
• Pixel statistics• Pixel intensities within the
image (histograms, moments)
The WNDCHRM Feature Set