incorporating artificial intelligence into mammography prediction louis oliphant computer sciences...

11
Incorporating Artificial Intelligence into Mammography Prediction Louis Oliphant Computer Sciences Department University of Wisconsin- Madison

Upload: jasper-shepherd

Post on 29-Jan-2016

225 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Incorporating Artificial Intelligence into Mammography Prediction Louis Oliphant Computer Sciences Department University of Wisconsin-Madison

Incorporating Artificial Intelligence into Mammography Prediction

Louis OliphantComputer Sciences DepartmentUniversity of Wisconsin-Madison

Page 2: Incorporating Artificial Intelligence into Mammography Prediction Louis Oliphant Computer Sciences Department University of Wisconsin-Madison

What is Artificial Intelligence?

The study and design of intelligent agents

Poole, Mackworth & Goebel 1998 http://www.bostondynamics.com/content/sec.php?section=BigDog GeneScan, C. Burge and S. Karlin 1997http://picasa.google.com/ http://www.toshiba.co.jp/about/press/2005_05/pr2001.htmhttp://babelfish.yahoo.com/

the spirit indeed is willing,but the flesh is weak.

De geest is wel gewillig,maar het vlees is zwak.

Page 3: Incorporating Artificial Intelligence into Mammography Prediction Louis Oliphant Computer Sciences Department University of Wisconsin-Madison

Y

0

0

1

0

0

0

Prediction:X

9 8 6 5 4 5 3 4 6

2 9 8 6 4 3 2 1 1

2 1 3 4 5 6 4 7 8

9 0 0 9 5 8 4 2 4

6 7 9 0 4 6 2 7 9

0 2 1 1 1 2 4 6 6

Prediction: Y

1

X

6 7 5 4 9 4 2 2 1

Supervised Machine Learning

X Y

1 3 2 4 1 1 8 3 3 0

2 3 2 4 6 2 1 7 2 0

4 5 6 7 7 7 2 1 7 1

4 5 6 2 1 7 8 2 6 0

3 2 1 1 1 0 4 3 3 1

5 4 3 1 6 4 7 3 2 0

7 8 6 7 5 3 4 1 7 0

5 6 7 7 4 2 3 1 1 1

0 9 8 9 7 4 6 3 2 0

Training Data

Classifier ModelTrained Classifier Model

Test DataY

0

0

1

0

0

1

Accuracy0.87

Nearest NeighborNeural NetworkSupport Vector Machine

X Y

1 3 2 4 1 1 8 3 3 0

2 3 2 4 6 2 1 7 2 0

4 5 6 7 7 7 2 1 7 1

4 5 6 2 1 7 8 2 6 0

3 2 1 1 1 0 4 3 3 1

5 4 3 1 6 4 7 3 2 0

7 8 6 7 5 3 4 1 7 0

5 6 7 7 4 2 3 1 1 1

0 9 8 9 7 4 6 3 2 0

Fixed Length Feature Vector

Page 4: Incorporating Artificial Intelligence into Mammography Prediction Louis Oliphant Computer Sciences Department University of Wisconsin-Madison

Fixed Length Feature Vector

First Order Logic Data

X y

1 3 2 4 1 1 8 3 3 0

2 3 2 4 6 2 1 7 2 0

4 5 6 7 7 7 2 1 7 1

4 5 6 2 1 7 8 2 6 0

3 2 1 1 1 0 4 3 3 1

5 4 3 1 6 4 7 3 2 0

7 8 6 7 5 3 4 1 7 0

5 6 7 7 4 2 3 1 1 1

0 9 8 9 7 4 6 3 2 0

Page 5: Incorporating Artificial Intelligence into Mammography Prediction Louis Oliphant Computer Sciences Department University of Wisconsin-Madison

Id Patient Date MassShape

… MassSize

Loc

1 P1 5/03 Oval 3mm RU4

2 P1 5/04 Round 8mm RU4

3 P2 5/04 Oval 4mm LL3

4 P3 6/00 Round 2mm RL2

… … … … … …

Mammography Dataset

Birads

3

5

1

4

Malignant/Benign

M

M

B

B

Collected from April 1999 to February 200418,270 Patients47,669 Mammograms510 Malignant 61,709 Benign

Page 6: Incorporating Artificial Intelligence into Mammography Prediction Louis Oliphant Computer Sciences Department University of Wisconsin-Madison

Inductive Logic Programming

Growth Medium: soil, woodCap Color: white, redGrouping: single, clusterAnnulus: present, not present

edible poisonous

edible(X) :- cap_color(X,red), annulus(X,present).

edible(X) :- medium(X,wood), grouping(X,single).

Page 7: Incorporating Artificial Intelligence into Mammography Prediction Louis Oliphant Computer Sciences Department University of Wisconsin-Madison

Our Model

Rules + TAN

Prolog + Java

Do: Find rule and add it if improves model performanceUntil time limit

Page 8: Incorporating Artificial Intelligence into Mammography Prediction Louis Oliphant Computer Sciences Department University of Wisconsin-Madison

Testing The Model

Purpose: assess future performance

Data set

Train set Test set

Score: 0.67

Small test setHigh variance

Cross Validation

Train setTest set

Score: 0.68

Score: 0.65

Score: 0.79

Score: 0.81

Average Score:0.72

Standard Deviation:0.07

Page 9: Incorporating Artificial Intelligence into Mammography Prediction Louis Oliphant Computer Sciences Department University of Wisconsin-Madison

Birads >= 5

Birads >= 4

Birads…

TAN

TAN + Rules

Our Results

Page 10: Incorporating Artificial Intelligence into Mammography Prediction Louis Oliphant Computer Sciences Department University of Wisconsin-Madison

Fielding The System

Approval process for pilot study Web page interface

XHTML, Javascript Enter Descriptors, Patient profile,

Radiologist’s Score Backend

Java + Prolog Return probability of malignancy and

Why model makes the prediction

Page 11: Incorporating Artificial Intelligence into Mammography Prediction Louis Oliphant Computer Sciences Department University of Wisconsin-Madison

Conclusions

Computers good at combining multiple interacting features

Adding rules improves performance Rules lend insight into predictive models

To Remember

Supervised Machine Learning Data set Model Evaluation Metric