mri-based biomarkers of therapeutic response in triple...
Post on 21-Jan-2021
1 Views
Preview:
TRANSCRIPT
MRI-Based Biomarkers of Therapeutic Responsein Triple-Negative Breast Cancer
Daniel Golden
Postdoctoral Scholar (Radiology)Stanford University
Daniel Rubin Laboratory
NCI Cancer Imaging Fellowship SeminarDecember 12, 2012
SCIT
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 1 / 19
Space Physics
Neutral Atmosphere
Advantage: Free Magnet! Disadvantage: Only 10-4 T
Radiation Belts
Satellites and
astronauts
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 2 / 19
Space Physics
Neutral Atmosphere
Advantage: Free Magnet!
Disadvantage: Only 10-4 T
Radiation Belts
Satellites and
astronauts
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 2 / 19
Space Physics
Neutral Atmosphere
Advantage: Free Magnet! Disadvantage: Only 10-4 T
Radiation Belts
Satellites and
astronauts
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 2 / 19
Space Physics
Neutral Atmosphere
Advantage: Free Magnet! Disadvantage: Only 10-4 T
Radiation Belts
Satellites and
astronauts
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 2 / 19
Space Physics
Neutral Atmosphere
Advantage: Free Magnet! Disadvantage: Only 10-4 T
Radiation Belts
Satellites and
astronauts
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 2 / 19
Motivation
Triple-Negative Breast Cancer15% of all breast cancers; 30,000 annual diagnoses; 8000 deathsLacks estrogen, progesterone, HER2 receptorsResponse to chemo is mixed
Critical NeedA way to predict in advance whether patients will respond:Precision Medicine
Treatment A
Treatment B
Known Malignancy Selection of Optimal Treatment
???
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 3 / 19
Motivation
Triple-Negative Breast Cancer15% of all breast cancers; 30,000 annual diagnoses; 8000 deathsLacks estrogen, progesterone, HER2 receptorsResponse to chemo is mixed
Critical NeedA way to predict in advance whether patients will respond:Precision Medicine
Treatment A
Treatment B
Known Malignancy Selection of Optimal Treatment
???
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 3 / 19
Motivation
Triple-Negative Breast Cancer15% of all breast cancers; 30,000 annual diagnoses; 8000 deathsLacks estrogen, progesterone, HER2 receptorsResponse to chemo is mixed
Critical NeedA way to predict in advance whether patients will respond:Precision Medicine
Treatment A
Treatment B
Known Malignancy Selection of Optimal Treatment
???
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 3 / 19
Motivation
Dynamic Contrast-Enhanced MRI Imaging (DCE-MRI)Acquires multiple images before and after contrast injectionWhole tumor, minimally-invasive (unlike biopsy)Reveals tumor kinetic phenotype: morphology and textureHypothesis: Features can predict treatment response
Treatment A
Treatment B
Known Malignancy Selection of Optimal Treatment
???Treatment A
Treatment B
Known Malignancy Selection of Optimal Treatment
???
MRI Features
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 4 / 19
Motivation
Dynamic Contrast-Enhanced MRI Imaging (DCE-MRI)Acquires multiple images before and after contrast injectionWhole tumor, minimally-invasive (unlike biopsy)Reveals tumor kinetic phenotype: morphology and textureHypothesis: Features can predict treatment response
Treatment A
Treatment B
Known Malignancy Selection of Optimal Treatment
???
Treatment A
Treatment B
Known Malignancy Selection of Optimal Treatment
???
MRI Features
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 4 / 19
Motivation
Dynamic Contrast-Enhanced MRI Imaging (DCE-MRI)Acquires multiple images before and after contrast injectionWhole tumor, minimally-invasive (unlike biopsy)Reveals tumor kinetic phenotype: morphology and textureHypothesis: Features can predict treatment response
Treatment A
Treatment B
Known Malignancy Selection of Optimal Treatment
???Treatment A
Treatment B
Known Malignancy Selection of Optimal Treatment
???
MRI Features
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 4 / 19
Data Set
The Triple-Negative Breast Cancer (TNBC) TrialClinical trial run by Melinda Telli and Jim Ford at Stanford93 patients with triple-negative or BRCA-mutated breast cancer69 patients available for analysisThis imaging study: retrospective and proof-of-concept
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 5 / 19
Example pre-chemo MRIs
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 6 / 19
Outline
1 Model Features
2 Modeling and Results
3 Conclusion and Future Work
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 7 / 19
Outline
1 Model Features
2 Modeling and Results
3 Conclusion and Future Work
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 8 / 19
List of Features
Semantic ImagingBreast Imaging Reporting and Data System BI-RADS
Quantitative ImagingLesion kinetic texture via the Gray-Level Co-Occurrence Matrix(GLCM)
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 9 / 19
Semantic Imaging Features
BI-RADSMass: shape, margins, enhancementNon-Mass: distribution, internal enhancement
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 10 / 19
Tumor Spatial Heterogeneity
Gene-expression signatures of good and poor prognosis were detected in different regions of the same tumor.
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 11 / 19
Tumor Spatial Heterogeneity
Gene-expression signatures of good and poor prognosis were detected in different regions of the same tumor.
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 11 / 19
Quantitative Imaging Features
The Gray-Level Co-Occurrence Matrix (GLCM)Based on texture of kinetic parameters
Lesion KineticImage
Pixel and Neighbor Values
2 4 6 8
2
4
6
8
2000
0
GLCM
Pix
el A
mpl
itude
Pixel Amplitude
Num
Pix
els
Countand Sum
Number of pixels with value 4 neighboring pixels with value 1
Contrast Correlation Energy HomogeneityScalar measures of image texture
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 12 / 19
Quantitative Imaging Features
The Gray-Level Co-Occurrence Matrix (GLCM)Based on texture of kinetic parameters
Lesion KineticImage
Pixel and Neighbor Values
2 4 6 8
2
4
6
8
2000
0
GLCM
Pix
el A
mpl
itude
Pixel Amplitude
Num
Pix
els
Countand Sum
Number of pixels with value 4 neighboring pixels with value 1
Contrast Correlation Energy HomogeneityScalar measures of image texture
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 12 / 19
Quantitative Imaging Features
The Gray-Level Co-Occurrence Matrix (GLCM)Based on texture of kinetic parameters
Lesion KineticImage
Pixel and Neighbor Values
2 4 6 8
2
4
6
8
2000
0
GLCM
Pix
el A
mpl
itude
Pixel Amplitude
Num
Pix
els
Countand Sum
Number of pixels with value 4 neighboring pixels with value 1
Contrast Correlation Energy HomogeneityScalar measures of image texture
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 12 / 19
Quantitative Imaging Features
The Gray-Level Co-Occurrence Matrix (GLCM)Based on texture of kinetic parameters
Lesion KineticImage
Pixel and Neighbor Values
2 4 6 8
2
4
6
8
2000
0
GLCM
Pix
el A
mpl
itude
Pixel Amplitude
Num
Pix
els
Countand Sum
Number of pixels with value 4 neighboring pixels with value 1
Contrast Correlation Energy HomogeneityScalar measures of image texture
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 12 / 19
Quantitative Imaging Features
The Gray-Level Co-Occurrence Matrix (GLCM)Based on texture of kinetic parameters
Lesion KineticImage
Pixel and Neighbor Values
2 4 6 8
2
4
6
8
2000
0
GLCM
Pix
el A
mpl
itude
Pixel Amplitude
Num
Pix
els
Countand Sum
Number of pixels with value 4 neighboring pixels with value 1
Contrast Correlation Energy HomogeneityScalar measures of image texture
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 12 / 19
Quantitative Imaging Features
The Gray-Level Co-Occurrence Matrix (GLCM)Based on texture of kinetic parameters
Lesion KineticImage
Pixel and Neighbor Values
2 4 6 8
2
4
6
8
2000
0
GLCM
Pix
el A
mpl
itude
Pixel Amplitude
Num
Pix
els
Countand Sum
Number of pixels with value 4 neighboring pixels with value 1
Contrast Correlation Energy HomogeneityScalar measures of image texture
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 12 / 19
Outline
1 Model Features
2 Modeling and Results
3 Conclusion and Future Work
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 13 / 19
Example Model Results
Modeling MethodologyLasso logistic regression(chooses optimalfeatures and createsregression model)Performance assessedvia cross-validated ROCcurves
This ModelResponse: residualtumor and lymph nodesFeatures: pre-chemotexture and BI-RADS
0 0.5 10
0.2
0.4
0.6
0.8
1
1 − Specificity
Sen
siti
vity
Predict Residual Nodes and Tumor
Good
Bad N=58
TextureAUC=0.5
BI-RADSAUC=0.77
Texture and BI-RADSAUC=0.88
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 14 / 19
Example Model Results
Modeling MethodologyLasso logistic regression(chooses optimalfeatures and createsregression model)Performance assessedvia cross-validated ROCcurves
This ModelResponse: residualtumor and lymph nodesFeatures: pre-chemotexture and BI-RADS
0 0.5 10
0.2
0.4
0.6
0.8
1
1 − Specificity
Sen
siti
vity
Predict Residual Nodes and Tumor
Good
Bad N=58
TextureAUC=0.5
BI-RADSAUC=0.77
Texture and BI-RADSAUC=0.88
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 14 / 19
Example Model Results
Modeling MethodologyLasso logistic regression(chooses optimalfeatures and createsregression model)Performance assessedvia cross-validated ROCcurves
This ModelResponse: residualtumor and lymph nodesFeatures: pre-chemotexture and BI-RADS
0 0.5 10
0.2
0.4
0.6
0.8
1
1 − Specificity
Sen
siti
vity
Predict Residual Nodes and Tumor
Good
Bad N=58
TextureAUC=0.5
BI-RADSAUC=0.77
Texture and BI-RADSAUC=0.88
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 14 / 19
Example Model Results
Modeling MethodologyLasso logistic regression(chooses optimalfeatures and createsregression model)Performance assessedvia cross-validated ROCcurves
This ModelResponse: residualtumor and lymph nodesFeatures: pre-chemotexture and BI-RADS
0 0.5 10
0.2
0.4
0.6
0.8
1
1 − Specificity
Sen
siti
vity
Predict Residual Nodes and Tumor
Good
Bad N=58
TextureAUC=0.5
BI-RADSAUC=0.77
Texture and BI-RADSAUC=0.88
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 14 / 19
Example Model Results
Modeling MethodologyLasso logistic regression(chooses optimalfeatures and createsregression model)Performance assessedvia cross-validated ROCcurves
This ModelResponse: residualtumor and lymph nodesFeatures: pre-chemotexture and BI-RADS
0 0.5 10
0.2
0.4
0.6
0.8
1
1 − Specificity
Sen
siti
vity
Predict Residual Nodes and Tumor
Good
Bad N=58
TextureAUC=0.5
BI-RADSAUC=0.77
Texture and BI-RADSAUC=0.88
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 14 / 19
Example Model Results
Modeling MethodologyLasso logistic regression(chooses optimalfeatures and createsregression model)Performance assessedvia cross-validated ROCcurves
This ModelResponse: residualtumor and lymph nodesFeatures: pre-chemotexture and BI-RADS
0 0.5 10
0.2
0.4
0.6
0.8
1
1 − Specificity
Sen
siti
vity
Predict Residual Nodes and Tumor
Good
Bad N=58
TextureAUC=0.5
BI-RADSAUC=0.77
Texture and BI-RADSAUC=0.88
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 14 / 19
Selected Features
−0.5 0 0.5 1
GLCM AUC energy
BI-RADS mass margin spiculated
BI-RADS non-mass
GLCM kep homogeneity
GLCM kep energy
BI-RADS mass shape round
BI-RADS mass enhancement homog.
feature weight
good responsepoor response
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 15 / 19
Outline
1 Model Features
2 Modeling and Results
3 Conclusion and Future Work
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 16 / 19
Conclusion and Future Work
ConclusionContrast-enhanced MRI can predict treatment responseBest model: combination of morphological and texture features
Future WorkImprove model
Extend to 3DNew quantitative features (e.g., region clustering via superpixels)Combine imaging with other biomarkers (e.g., genomics)
Try to predict survival
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 17 / 19
Conclusion and Future Work
ConclusionContrast-enhanced MRI can predict treatment responseBest model: combination of morphological and texture features
Future WorkImprove model
Extend to 3DNew quantitative features (e.g., region clustering via superpixels)Combine imaging with other biomarkers (e.g., genomics)
Try to predict survival
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 17 / 19
Thank You
MentorDaniel Rubin
CollaboratorsJafi LipsonMelinda TelliJim FordKatie PlaneyNick Hughes
FundingStanford SCIT Program (NIH T32 CA009695)NIH U01 CA142555
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 18 / 19
Non-Imaging Features
ClinicalAge at diagnosisTumor stage (IA–IIIA)Tumor grade (II or III)T and N stage from TNM (T0–T4, N0–N3)ER/PR percent (for non-triple-negative)Ki67 percentCycles of treatment received (4 or 6)
GenomicBRCA 1/2 mutation status
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 19 / 19
Other Models Features
BI−RADSGLCM Pre and BI−RADS
Patterns of ResponseGLCM Pre and GLCM Post
GLCM PreKi67
All Clinical but Ki67All Clinical
Feature Sets Lasso Model Response
Residual Tumor
Residual Lymph Nodes
Residual Tumorand Nodes
GLCM Post
Best Models
Residual tumor and nodes:imaging (as shown)Residual tumor: post-chemo
textureResidual nodes: clinical(imaging good too)
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 20 / 19
Tumor/Nodes Model Results
SensitivitySpecificityAUC
junkjunk
junk
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Predict residual nodes
junk
junkjunk
junk
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Predict residual tumor and nodes
junkjunkjunk
junk
junkjunk
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
GLCM Pre and BI−RADSBI−RADS
Patterns of ResponseGLCM Pre and GLCM Post
GLCM PostGLCM Pre
Ki67All Clinical but Ki67
All Clinical
Predict residual tumor
n=37 n=51 n=44 n=51 n=44 n=41 n=55 n=58 n=51
GLCM Pre and BI−RADSBI−RADS
Patterns of ResponseGLCM Pre and GLCM Post
GLCM PostGLCM Pre
Ki67All Clinical but Ki67
All Clinical
junk ≡ AUC < 0.6
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 21 / 19
Residual Tumor Model Features
Relevant features forpredicting No Residual Tumor
0.5 0.75 1AUC
Ki67
−0.5 0 0.5
ki67 percent
b*std
0.5 0.75 1AUC
GLCM Post−chemotherapy
−1 0 1
avg wash out post−chemoGLCM wash out slope contrast post−chemo
GLCM Ktrans correlation post−chemo
b*std
0.5 0.75 1AUC
GLCM Pre− and GLCM Post−chemotherapy
−1 0 1
avg wash out post−chemoGLCM wash out slope contrast post−chemo
GLCM kep contrast post−chemoGLCM AUC contrast pre−chemo
GLCM Ktrans correlation post−chemo
b*std
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 22 / 19
Residual Nodes Model FeaturesRelevant features for
predicting No Residual Lymph Nodes
0.5 0.75 1AUC
All Clinical
−1 0 1
TNM N0TNM N1
stage IIIA4 treatment cycles
BRCA1 result negative
0.5 0.75 1AUC
All Clinical but Ki67
−1 0 1
TNM N0BRCA2 result negative
age at diagnosisTNM N1
ER status 1+4 treatment cycles
stage IIIABRCA1 result negative
0.5 0.75 1AUC
GLCM Pre−chemotherapy
−0.5 0 0.5
GLCM kep contrast pre−chemoGLCM AUC homogeneity pre−chemo
GLCM kep homogeneity pre−chemo
b*std
0.5 0.75 1AUC
GLCM Post−chemotherapy
−0.5 0 0.5
lesion area post−chemoGLCM Ktrans correlation post−chemo
b*std
0.5 0.75 1AUC
BI−RADS
−1 0 1
BI−RADS mass shape roundBI−RADS mass margin smooth
BI−RADS mass margin spiculatedBI−RADS non−mass−like
b*std
0.5 0.75 1AUC
GLCM Pre−chemotherapy and BI−RADS
−0.5 0 0.5
BI−RADS mass shape roundGLCM kep contrast pre−chemoGLCM kep energy pre−chemo
BI−RADS non−mass−likeGLCM AUC energy pre−chemo
GLCM kep homogeneity pre−chemo
b*std
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 23 / 19
Residual Tumor and Nodes Model Features
Relevant features forpredicting Residual Tumor and Lymph Nodes
0.5 0.75 1AUC
All Clinical but Ki67
−1 0 1
4 treatment cyclesER status 1+
tumor grade IIBRCA1 result negative
stage IIIATNM N1TNM T1TNM N2TNM N0
b*std
0.5 0.75 1AUC
GLCM Post−chemotherapy
−1 0 1
GLCM Ktrans correlation post−chemolesion area post−chemo
b*std
0.5 0.75 1AUC
GLCM Pre− and GLCM Post−chemotherapy
−0.5 0 0.5
GLCM Ktrans correlation post−chemolesion area post−chemo
b*std
0.5 0.75 1AUC
BI−RADS
−1 0 1
BI−RADS non−mass−likeBI−RADS mass margin spiculated
BI−RADS mass enhancement homogeneousBI−RADS mass shape round
b*std
0.5 0.75 1AUC
GLCM Pre−chemotherapy and BI−RADS
−1 0 1
GLCM AUC energy pre−chemoBI−RADS mass margin spiculated
BI−RADS non−mass−likeGLCM kep homogeneity pre−chemo
GLCM kep energy pre−chemoBI−RADS mass shape round
BI−RADS mass enhancement homogeneous
b*std
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 24 / 19
Dynamic Contrast-enhanced MRI
0 50 100 150 200
400
600
800
1000
Time (sec)
Avg
voxe
lint
ensi
ty
t = 0 sec
1 cm
t = 51 sec t = 61 sec t = 195 sec
Wash In ≈ Ktrans
Wash Out ≈ kep
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 25 / 19
Kinetic Modeling
t=1.5 min
1 cm
0 5 10 15−1
0
1
2
3
Minutes
Fractionalenhancement
DataModel
0
1
2
3ve(unitless)
0.5
1
1.5
2
Ktrans (min−1)
0.5
1
1.5
2
Kep(min−1)
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 26 / 19
Kinetic Modeling
t=1.5 min
1 cm 0 5 10 15−1
0
1
2
3
Minutes
Fractionalenhancement
DataModel
0
1
2
3ve(unitless)
0.5
1
1.5
2
Ktrans (min−1)
0.5
1
1.5
2
Kep(min−1)
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 26 / 19
Kinetic Modeling
t=1.5 min
1 cm 0 5 10 15−1
0
1
2
3
Minutes
Fractionalenhancement
DataModel
0
1
2
3ve(unitless)
0.5
1
1.5
2
Ktrans (min−1)
0.5
1
1.5
2
Kep(min−1)
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 26 / 19
Breast DCE-MRI Heterogeneity Review
Malignancy Survival Type TreatmentResponse
Texture
Sinha et al., 1997; Chen etal., 2007; Woods et al.,
2007; Kale et al., 2008; Nieet al., 2008; Agner et al.,2011; Karahaliou et al.,
2012
Holli etal.,
2010
Histogram Hauth et al., 2008; Preim etal., 2011
Johansenet al.,2009
Chang etal., 2004;
Padhani etal., 2009
Caveats
Generally considered all BC subtypes together
Usually reported simple t-tests for each feature; lacked multivariate regressionand cross-validation
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 27 / 19
Breast DCE-MRI Heterogeneity Review
Malignancy Survival Type TreatmentResponse
Texture
Sinha et al., 1997; Chen etal., 2007; Woods et al.,
2007; Kale et al., 2008; Nieet al., 2008; Agner et al.,2011; Karahaliou et al.,
2012
Holli etal.,
2010
YouAre
Here
Histogram Hauth et al., 2008; Preim etal., 2011
Johansenet al.,2009
Chang etal., 2004;
Padhani etal., 2009
Caveats
Generally considered all BC subtypes together
Usually reported simple t-tests for each feature; lacked multivariate regressionand cross-validation
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 27 / 19
Residual Cancer Burden
1
1
2
2
2
3
3
3
34
4
4
5
5
Term 1 (Primary Tumor)
Ter
m2
(Pos
itiv
eN
odes
)
0 1 2 30
0.5
1
1.5
2
2.5
3
RC
B
0
1
2
3
4
5
6
0<RCB<2.5
(26, 46%)RCB=0
(pCR)
(19, 33%)
RCB>2.5
(12, 21%)
Residual
Nodes
(3, 5%)
pCR
(19, 33%)Residual
Tumor
(22, 39%)Tumor
+ Nodes
(13, 23%)
0 1 2 3 4 50
0.2
0.4
0.6
0.8
1
RCB Value
Cum
ulat
ive
Dis
trib
utio
nF
un
ctio
n
Natural separation points
Only 1 case with residual tumor and nodes and RCB<2.5
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 28 / 19
RCB Model Results
SensitivitySpecificityAUC
junk
junkjunkjunk
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
b) Predict RCB > 2.5
junkjunk
junkjunkjunk
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
GLCM Pre and BI−RADSBI−RADS
Patterns of ResponseGLCM Pre and GLCM Post
GLCM PostGLCM Pre
Ki67All Clinical but Ki67
All Clinical
a) Predict pCR (RCB=0)
n=39 n=53 n=47 n=54 n=44 n=41 n=60 n=64 n=54
Daniel Golden (dgolden1@stanford.edu) MRI-Based Biomarkers Dec 12, 2012 29 / 19
top related