UroComp Home
Go to UroComp home

Prostate Cancer Detection Neural Network

Portions of this document were adapted in part from our paper:

Kalra P, Togami J, Bansal G, Partin AW, Brawer MK, Babaian RJ, Ross LS, Niederberger CS: A Neurocomputational Model for Prostate Carcinoma Detection: Cancer, Volume 98, Issue 9 (1 November 2003) .

The reader is strongly encouraged to read this paper prior to using the network, as it contains a more specific description of the model.


Contents


About Prostate Cancer Detection

Prostate cancer is the leading form of cancer in American men. Detection of prostate cancer has dramatically improved since the advent of prostate specific antigen (PSA), a blood test. This marker in itself, however, is limited by its lack of specificity. In other words, there are other benign conditions, such as prostatitis or BPH ("benign prostatic hyperplasia", a benign growth of the prostate which commonly occurs with age), which may falsely elevate a man's PSA. With the current accepted PSA threshold of 4.0 ng/mL, only 25% of men undergoing a biopsy for a PSA between 4.0 and 10.0 ng/mL actually receive a diagnosis of prostate carcinoma.

In clinical practice, physicians usually assess a patient's risk for having prostate cancer using information in addition to just the individual's PSA. There are several well described risk factors for prostate cancer, such as African American descent, or a strong family history of prostate cancer. These characteristics also need to be considering when counseling patients. For example, an African American man who has several family members with a history of prostate cancer may be advised to undergo a prostate biopsy even if his PSA is 3.9 ng/mL. This recommendation may be further reinforced by informing the patient of his relative risk for prostate cancer, which may be calculated using a neural network.

Return to top of document


Neural Network Programming and Training

3,268 men recently evaluated for early detection of prostate cancer served as the database for the study. The study population consisting of 354 subjects with known prostate biopsy outcomes was used to construct the dataset. Patients with PIN were excluded from this study. The resulting dataset of 348 subjects was modeled using "neUROn2++", a suite of C++ programs we designed to implement neural computational and statistical algorithms, which are cross compiled using Microsoft Visual C++ version 6 and GNU C++ (Cygwin port version 2.95).

The 7 clinical features evaluated include age, race, family history, International Prostate Symptom Score (IPSS), digital rectal examination (DRE), and total and complexed PSA (Bayer Diagnostics), and were encoded as the input nodes to the neural network. Subjects with PSA < 4.0 ng/ml represented approximately 20.7% of the dataset. The output node represented the result of the prostate biopsy, which was encoded as a binary variable (cancer or benign).

The dataset was randomly divided into a training set of 218 subjects, and a test set of 144 subjects. The proportion of positive biopsy outcomes was constrained to be similar in both sets using a randomization algorithm. The test set was excluded from training, and only used for cross-validation (n1/n2 method). A 1-hidden node layer with 3 nodes was determined to represent an optimal network architecture which maintained acceptable goodness-of-fit without overlearning. The training method was canonical off-line backpropagation with weight decay, with the weight decay term lambda chosen to be 5e-05.1 The network to be trained to completion when the error was observed to be oscillating at a local error minimum.

Wilk's Generalized Likelihood Ratio Test (GLRT) was employed to determine which input features were significant to the model's outcome in a reverse regression analysis.1 All input variates were found to be highly significant to the model (p << 0.000001). The dataset was also modeled using logistic regression and linear and quadratic discriminant function analysis (LDFA and QDFA) to compare the nonlinear computational method of neural computation with traditional linear statistical modeling tools.

1A description of this method of training, including weight decay and feature extraction using Wilk's GLRT, may be found in Golden RM, Mathematical methods for neural network analysis and design, Cambridge, MA: MIT Press, 1996.

Return to top of document


Accuracy of the Neural Network Compared to Linear Methods

Method
ROC Area2 p-value3
Neural Network
0.825
---
Logistic Regression
0.510
0.0001
Total PSA alone
0.678
0.0007
Complexed PSA alone
0.697
0.002
Quadratic Discriminant Function Analysis
0.011
0.0001
Linear Discriminant Function Analysis
0.674
0.0002

2Receiver Operating Characteristic Curve area. Numbers approach 1.0 as accuracy improves (a value of 1.0 would indicate sensitivity and specificity both 1.0). Areas were computed using the statistical method described by Wickens: Wickens TD, Elementary signal detection theory, New York: Oxford University Press, 2002.

3p-value indicates significance of difference from neural computational model, computed by DeLong's test: DeLong ER, DeLong DM, Clarke-Pearson DL: Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach, Biometrics1988; 44: 837-845.

Return to top of document


Click here to download psanet.prc, a PalmOS application for the PalmPilot or Handspring Visor device. (If you are using Netscape Navigator, and have trouble with this link, try clicking on it with the right mouse button, and choosing "Save Link As..." from the pop-up menu.)

  

Return to top of document