Predicting Breast Cancer

Predicting
Breast Cancer

Max Mitchell
Amanda Rollason
Cheannette Smith

AGENDA
 Understanding the Problem
 Exploring the Data
 Frequency Tables
 Graphs
 Modeling with Logistic Regression
 Validating the Predictive Model

THE PROBLEM
Of women who have a mammogram interpretation
that leads to a breast biopsy, 70% actually have
benign outcomes, which means only 30% sensitivity.

Understand Explore Model Validate

A SOLUTION
Better diagnostic tools for physicians would help
increase the sensitivity and specificity of
mammogram interpretations, while reducing the
number of unnecessary breast biopsies.


THE DATA SOURCE
 Dr. Rudiger Schulz-Wendtland, M. Elter, and T.
Wittenberg at the Institute of Radiology at University
of Erlangen in Nuremberg, Germany
 Data collected between 2003-2006
 Published in Medical Physics in 2007: “The Prediction
of Breast Cancer Biopsy Outcomes Using Two CAD
Approaches that both Emphasize an Intelligible
Decision Process”
 Donated database to University of California-Irvine
for their machine learning database repository


THE CASE-CONTROL STUDY
 961 observations represent the outcome for
women who already had breast biopsies
 516 benign cases
 445 malignant cases
 Two physicians reviewed the mammograms
 not knowing that each woman already had a biopsy
for the suspected mass
 not knowing the outcome of her biopsy
 Computer aided-diagnosis (CAD) systems were
run for the full-field digital mammograms

THE VARIABLES
 1 binary response, the SEVERITY of mass lesion
 Malignant = 1
 Benign = 0
 1 continuous predictor, AGE
 1 semi-ordinal predictor, the BI-RADS Assessment
 3 nominal predictors, the CAD output
 SHAPE
 DENSITY
 MARGIN

THE MAMMOGRAM DATA


Breast Imaging Reporting and Data System
BI-RADS
ASSESSMENT CLINICAL RECOMMENDATION
CATEGORY
Assessment Need to review prior studies
0
Incomplete and/or complete additional imaging

1 Negative Continue routine screening

2 Benign Continue routine screening

Probably Benign Short-term follow-up at 6 months,
3
Finding then every 6 to 12 months for 1 to 2 years

4 Suspicious Abnormality Perform biopsy, preferably needle biopsy

Highly Suspicious of
5 Biopsy and treatment, as necessary
Malignancy
Known Biopsy–
6 Assure that treatment is completed
Proven Malignancy

SHAPE MARGIN
1 = ROUND 1 = CIRCUMSCRIBED

2 = OVAL 2 = MICROLOBULATED

3 = LOBULAR 3 = OBSCURED

4 = IRREGULAR 4 = ILL-DEFINED

5 = SPICULATED


DENSITY
1 = high 2 = iso

4=fat-
containing
3=low


EXPLORING THE DATA


LOGISTIC REGRESSION MODELS
 There is a binary response variable.
 There are more than three predictors, so
frequency tables alone will be inadequate.
 The predictors are both numerical and
categorical.
 Some of the categorical variables are ordinal.


PRE-MODELING
MODEL MAIN EFFECTS – Mostly Categorial
1 AGE BIRADSc SHAPEc MARGINc DENSITYc
2 AGE BIRADSc
3 AGE SHAPEc MARGINc DENSITYc
4 AGE SHAPEc MARGINc
5 AGE BIRADSc SHAPEc MARGINc
6 AGE BIRADSc SHAPEc


PRE-MODELING
MODEL MAIN EFFECTS – Mostly Numerical
7 AGE BIRADS SHAPE MARGIN DENSITYc
8 AGE BIRADS
9 AGE SHAPE MARGIN DENSITYc
10 AGE SHAPE MARGIN
11 AGE BIRADS SHAPE MARGIN
12 AGE BIRADS SHAPE


MODELING THE DATA
MODEL TERMS

A AGE BIRADSc SHAPEc MARGINc DENSITYc

B AGE BIRADSc SHAPEc MARGINc

C AGE BIRADSc SHAPEc

D AGE BIRADS SHAPE

E AGE BIRADS SHAPE AGE×BIRADS SHAPE×BIRADS


LOGISTIC REGRESSION – Training
MODEL p-cutoff Sensitivity Specificity AIC AUC = c

A B c S c Mc D c
0.414 0.845 0.838 475.3 0.910
QCS

A Bc S c M c 0.363 0.878 0.815 507.6 0.906

A Bc Sc 0.364 0.875 0.815 534.7 0.902

ABS 0.419 0.844 0.813 563.4 0.888

A B S AB SB 0.438 0.873 0.809 544.0 0.899


LOGISTIC REGRESSION
Logit ( ˆ) α β1 AGE β2BIRADS β3SHAPE
β4 AGE BIRADS β5BIRADS SHAPE


LOGISTIC REGRESSION – Validation
MODEL Sensitivity Specificity AUC 95% CI

A Bc S c M c D c
0.866 0.863 (0.822, 0.907)
QCS

A Bc S c Mc 0.858 0.850 (0.812 ,0.897)

A Bc S c 0.846 0.844 (0.802, 0.887)

ABS 0.835 0.848 (0.798, 0.884)

A B S AB SB 0.816 0.870 (0.800, 0.928)


Comparing ROC curves


Example 1
AGE = 42 | BIRADS = 2 | SHAPE = Oval = 2

Logit = -34.1514 + 0.2398(42) + 6.8365(2) + 4.0423 (2)
– 0.0441(42)(2) – 0.7842(2)(2)
= -8.903
Odds = e-8.903 = 0.0001

 Patient most likely does not have a malignant lesion.

TRUE. She had multiple cutaneous neurofibromas. They
are benign, so there is no evidence of malignancy. The
reader recommended that she should have a normal
interval screening follow-up in 12 months.


Example 2
AGE = 62 | BIRADS = 4 | SHAPE = Irregular = 4

Logit = -34.1514 + 0.2398(62) + 6.8365(4) + 4.0423 (4)
– 0.0441(62)(4) – 0.7842(4)(4)
= 1.5162
Odds = e1.5162 = 4.55

 Patient most likely does have a malignant lesion.

TRUE. She had invasive ductal carcinoma, so there was
evidence of malignancy. The reader saw that she had a
suspicious abnormality and recommended a core needle
biopsy.


Conclusion
 Readers’ interpretation alone (BIRADS) isn’t
sufficient.
 Computer Aided-Diagnosis systems (SHAPE,
MARGIN, and DENSITY) alone aren’t sufficient.
 AGE does need to be considered for determining if
a breast biopsy is warranted.
 AGE, BIRADS, and SHAPE did the most to improve
sensitivity and specificity.


For the Future
 Incorporate other CAD tools.
 MRI tests
 Ultrasound examinations
 Explore results of other modeling methods.
 Decision Trees
 Boot-strapping
 Educate patients regarding the imperfect process
of mammogram interpretation.


Questions?


Predicting Breast Cancer

Recomendados

Recomendados

Mais conteúdo relacionado

Semelhante a Predicting Breast Cancer

Semelhante a Predicting Breast Cancer (20)

Predicting Breast Cancer

Notas do Editor