Using Artificial Neural Using Artificial Neural
Networks to
Networks to Predict Predict
Malignancy of Ovarian Tumors Malignancy of Ovarian Tumors
C. Lu1, J. De Brabanter1, S. Van Huffel1, I. Vergote2, D. Timmerman2
1Department of Electrical Engineering,
Katholieke Universiteit Leuven, Leuven, Belgium, Department of Obstetrics and Gynecology,
Overview Overview
Introduction
Data Exploration
Input Selection
Model Building
Model Evaluation
Conclusions
Introduction Introduction
Problem
ovarian masses: a common problem in gynecology.
develop a reliable diagnostic tool to discriminate
preoperatively between benign and malignant tumors.
assist clinicians in choosing the appropriate treatment.
Data
Patient data collected at Univ. Hospitals Leuven, Belgium, 1994~1999
425 records, 25 features.
291 benign tumors, 134 (32%) malignant tumors.
Introduction Introduction
Methods
Data exploration:
Data preprocessing, univariate analysis, PCA, factor analysis, discriminant analysis, logistic regression…
Modeling:
Logistic regression (LR) models
Artificial neural networks (ANN): MLP, RBF
Performance measures:
Receiver operating characteristic (ROC) analysis
ROC curves
constructed by plotting the sensitivity versus the 1- specificity, or false positive rate, for varying probability cutoff level.
visualization of the relationship between
sensitivity and specificity of a test.
Area under the ROC curves (AUC)
measures the probability of the classifier to correctly classify events and
nonevents.
Data exploration Data exploration
Univariate analysis:
preprocessing:
descriptive statistics, histograms…
Variable (symbol) Benign Malignant
Demographic Age (age)
Postmenopausal (meno) 45.6 15.2
31.0 % 56.9 14.6
66.0 %
Serum marker CA 125 (log) (l_ca125) 3.0 1.2 5.2 1.5
CDI High color score (colsc3,4) 19.0% 77.3 %
Morphologic Abdominal fluid (asc) Bilateral mass (bilat) Unilocular cyst (un)
Multiloc/solid cyst (mulsol) Solid (sol)
Smooth wall (smooth) Irregular wall (irreg) Papillations (pap)
32.7 % 13.3 % 45.8 % 10.7 % 8.3 % 56.8 % 33.8 % 12.5 %
67.3 % 39.0 % 5.0 % 36.2 % 37.6 % 5.7 % 73.2 % 53.2 %
Demographic, serum marker, color Doppler imaging and morphologic variables
Data exploration Data exploration
Multivariate analysis:
factor analysis
biplots
Fig. Biplot of Ovarian Tumor data.
The observations are plotted as points (0=benign,
1=malignant), the variables are plotted as vectors from the origin.
- visualization of the correlation between the variables
- visualization of the relations between the variables and clusters.
Input Selection Input Selection
Stepwise logistic regression analysis
Searching in the feature space
fix several of the most significant variables, then vary combinations with the other predictive variables.
different logistic regression models with different subsets of input variables were built and validated.
subsets of variables were selected according to their predictive performance on the training set and test set.
Model building Model building
Logistic regression (LR) model
Artificial neural networks
feed-forward neural networks, universal approximators:
- multi-layer perceptron (MLP)
- generalized regression network (GRNN)
generalization capacity: central issue during network design and training.
Model building Model building
- - LR LR
Parameter estimation:
- maximum likelihood - iterative procedure
. . . . .
b i a s
P r o b a b i l i t y o f m a l i g n a n c y
g
) exp(
1 ) 1
(a a
g
Fig. Architecture of LRs for Predicting Malignancy of Ovarian Tumors
structure:
Training
Bayesian regularization combined with Levenberg- Marquardt optimization.
Model Building Model Building
- ANN - MLP - ANN - MLP
M : n u m b e r o f h i d d e n n e u r o n s
d : n u m b e r o f i n p u t v a r i a b l e s
. . . . .
b i a s
b i a s
P r o b a b i l i t y o f m a l i g n a n c y
g
g
g
g
M O D E L 1 : m e n o c o l s c 3 c o l s c 4 l _ c a 1 2 5 a s c s o l i r r e g p a p M O D E L 2 : m e n o c o l s c 3 c o l s c 4 l _ c a 1 2 5 a s c s m o o t h p a p
) exp(
1 ) 1
( a a
g
M j
d i
i ji
j g w x
w g
y
0 0
) 1(
) 2 (
Fig. Architecture of MLPs for Predicting Malignancy of Ovarian Tumors
structure
MLP1: 8-3-1 MLP2: 7-3-1
Model Building Model Building
– ANN - GRNN – ANN - GRNN
Fig. Architecture of GRNNs for Predicting Malignancy of Ovarian Tumors
. . . . . o u t p u t
N
j j N j
j
x x t
x y
j
1 1
) (
) ( )
(
ti: t a r g e t o u t p u t o f i t h t r a i n i n g d a t a
2
2
exp 2 ) (
j i
j h
x x x
N : # t r a i n i n g d a t a x : i n p u t v e c t o r
x : i n p u t v e c t o r g
2
1 N
… … …
… … …
t1 t2 tN
Training:
GRNN is another term for Nadaraya- Watson kernel regression. No iterative training; the widths of RBF units h act as smoothing parameters, chosen by cross- validation.
structure
•RMI: risk of malignancy index = score ×
Training set : data from the first treated 265 patients
Test set : data from the latest treated 160 patients
Model Evaluation Model Evaluation
- Holdout CV - Holdout CV
AUC estimates and standard errors from hold out CV
stratified 7-fold CV
for each run of 7- fold CV:
mAUC :
(iAUCi)/7, i =1,…7, AUCi is the AUC on the ith validation set
expected ROC:
Averaging.
Repeat 7-fold CV 30 times with different partitions => better
Model Evaluation Model Evaluation
- K-fold CV - K-fold CV
Box plot of meanAUC from 7-fold CVExpected ROC curves from k-fold CV
Multiple comparison of mAUCs:
one-way ANOVA followed by Tukey multiple comparison
.
Rank ordered significant subgroups from multiple comparison on mean AUC
Models RMI LR2 LR1 GRN1 GRN2 MLP2 MLP1
mean
mAUC 0.882 0.943 0.954
SD 0.003
0.939 0.003
0.941
0.004 0.003
0.944 0.003
0.944
0.003 0.003
Note: The subsets of adjacent means that are not significantly different at 95% confidence level are indicated by drawing a line under the subsets.
Model Evaluation Model Evaluation
- K-fold CV
- K-fold CV
Conclusions Conclusions
Summary
AUC is the advocated performance measure
Data exploratory analysis helps to analyze the data set.
MLPs have the potential to give more reliable prediction.
Future work
Develop models with kernel methods, e.g. LS-SVM
ANNs are blackbox models. A hybrid methodology, greybox models might be more promising