1 of 21

Detecting Prostate Cancer Using MRI Data

David Anderson, Bruce Golden, Ed Wasil, Howard Zhang

1

INFORMS Annual Meeting October, 2013

2 of 21

Prostate Cancer

  • The NCI estimates that 15% of men born today will be diagnosed with prostate cancer

​

  • Average costs of $10,000 in the first year after diagnosis

​

  • Hard to diagnose

2

3 of 21

Prostate Cancer Diagnosis Methods

  • PSA Test
    • Non-intrusive
    • High false positive rate
      • 67% sensitivity, 58% specificity (Thompson et al. 2005)
  • Digital Exam
    • Inconsistent
  • Biopsy
    • Painful
    • Expensive
    • Possibly severe side effects

3

4 of 21

MRIs to the Rescue?

4

5 of 21

Research Question

  • Can we use MRIs to screen for prostate cancer?

​

    • Will doing so be more cost effective than the current system?

5

6 of 21

Data

  • 223 slices of prostates from radical prostatectomy patients

​

  • 3 types of MRIs on each slice (Dynamic Contrast Enhanced, Diffusion Weighted, and Magnetic Resonance Spectroscopic Imaging)

​

  • 119 had cancer (Gleason score of 5 or above)

​

6

7 of 21

Independence of Slices

  • Slices from the same prostate may have similar cancer status and MRI data
  • Correlation between slices of the same prostate would bias our performance upwards
  • Correlation in Gleason scores of adjacent slices is 0.30, and for slices two apart it is 0.004

7

8 of 21

Distribution of Cancer

8

9 of 21

Three Methods

  • Logistic Regression

​

  • Nearest Neighbors Clustering

​

  • Augmented Logistic Regression

9

10 of 21

Results – Logistic Regression

10

11 of 21

Results – Nearest Neighbors

11

12 of 21

Augmented Logistic Regression Results

12

13 of 21

Augmented Logistic Regression Results

13

Β 

Gleason Score

Β 

0 – 4

5 – 8

Predicted Healthy

79

22

Predicted Cancer

25

97

The combined model achieves 82% sensitivity and 76% specificity

14 of 21

High Severity Cancer

  • Many prostate cancers are slow growing
    • β€œMore men die β€˜with’ prostate cancer than β€˜from’ it”

​

  • Identifying high severity cancer (scores of 7 or 8) is important

​

​

14

15 of 21

High Severity Results

15

16 of 21

High Severity Results

16

Β 

Gleason Score

Β 

0 – 6

7 – 8

Predicted Healthy

151

5

Predicted Cancer

36

31

For high severity cancers, the combined model achieves 81% sensitivity and 86% specificity

17 of 21

Cost Effectiveness

  • Prices for medical services vary widely
    • Biopsies average ~$2100
    • MRIs average ~$700

​

  • If MRIs can reduce the number of biopsies by at least 1/3 they will reduce costs

​

​

​

17

18 of 21

Conclusions

  • MRIs can be used to identify prostate cancer

​

  • By looking at each slice of a prostate we can identify where to biopsy

​

  • MRIs offer possibly better predictive power than PSA tests, and are less invasive than biopsies

​

18

19 of 21

Contribution

  • Combine MRI types

​

  • Automated prediction

​

  • Distinguish between high and medium severity cancers

​

​

19

20 of 21

Future Work

  • Collect more data
    • Healthy patients and cancerous

​

  • Build models for whole prostates, not slices

​

  • Predict specific Gleason scores

​

​

20

21 of 21

Questions?

David.Anderson@Baruch.CUNY.edu

21