A.N.N.C.R.I.P.SA.N.N.C.R.I.P.SThe Artificial Neural Networks The Artificial Neural Networks

forfor

Cancer Research in Prediction Cancer Research in Prediction & Survival & Survival

A CSI – VESIT PRESENTATION

Presented By

Karan Kamdar Amit Mathapati Arpan Nanavati

© 2003 – The A.N.N.C.R.I.P.S

Why A.N.N.C.R.I.P.S ?Why A.N.N.C.R.I.P.S ? Prostate Cancer – MostProstate Cancer – Most common form among mencommon form among men

ScreeningScreening – Carried out through blood tests & presence – Carried out through blood tests & presence of high PSA (can indicate cancer risk) of high PSA (can indicate cancer risk)

DrawbacksDrawbacks – Initial screenings lead to high percentage – Initial screenings lead to high percentage of false positive test results (FPTRs) of false positive test results (FPTRs)

Scope for improvementScope for improvement – FPTRs can be reduced by – FPTRs can be reduced by employing intelligent Artificial Neural Networks employing intelligent Artificial Neural Networks

Implementation at all levelsImplementation at all levels – ANN models can be used at every stage of – ANN models can be used at every stage of Cancer analysis Cancer analysis

AdvantageAdvantage – Non-required trial and error methods reduced. Hence early – Non-required trial and error methods reduced. Hence early cancer diagnosis and treatment cancer diagnosis and treatment

Project OverviewProject Overview

Voluntary effort started by the core members of this group and our colleaguesVoluntary effort started by the core members of this group and our colleagues

Objective – Build a mathematical model to improve prostate cancer detection and staging Objective – Build a mathematical model to improve prostate cancer detection and staging systemssystems

Basis – Mathematical model revolves around the concept of Artificial Neural Networks (ANNs)Basis – Mathematical model revolves around the concept of Artificial Neural Networks (ANNs)

Development & Implementation – Programming the ANN model to develop a standalone Development & Implementation – Programming the ANN model to develop a standalone application which can be deployed across medical organizations application which can be deployed across medical organizations

Introduction to ANNsIntroduction to ANNs Information processing paradigm inspired by the way biological nervous systems eg. brain process information

ANNs adopt this interconnected neuron network to perform complex computations

Basic ANN modelBasic ANN model

Overview of Prostate CancerOverview of Prostate Cancer

Prostate – Male sex gland producing the semenProstate – Male sex gland producing the semen

Prostate Cancer –– Cancer beginning in prostate which may remain in the prostate gland or spread

Screening & Diagnosis –– If symptoms occur, DRE & blood tests are undertaken to measure level of PSA

Staging –– After initial diagnosis, further staging such as TNM undertaken

Our ANN model employed to improve cancer risk estimation of screening & staging systems

Linking ANNs to PCa AnalysisLinking ANNs to PCa Analysis

Finding % tPSA – Current method of PCa detectionFinding % tPSA – Current method of PCa detection

Faults – 1. Only Cancer confined to the prostate detectedFaults – 1. Only Cancer confined to the prostate detected

2. Additional info requires more tissue samplings 2. Additional info requires more tissue samplings

Possible Soln. – Finding % free PSA (fPSA) to suggest Possible Soln. – Finding % free PSA (fPSA) to suggest

spread of Cancer to other parts. spread of Cancer to other parts.

Core Drawbacks – Both systems not intelligent. Core Drawbacks – Both systems not intelligent.

a) Cannot predict on individual basis a) Cannot predict on individual basis

b) Inability to learn and associate b) Inability to learn and associate

c) Moderate Accuracy c) Moderate Accuracy

The A.N.N.C.R.I.P.S modelThe A.N.N.C.R.I.P.S model – Take the best of both – Take the best of both

and include other role playing variables to build an intelligent,and include other role playing variables to build an intelligent,

reliable cancer detection and prediction system reliable cancer detection and prediction system

Building the ANN modelBuilding the ANN model

Based on the concept of MultiLayer Perceptron (MLP)Based on the concept of MultiLayer Perceptron (MLP)

Consists of a network of processing elements or nodes arranged in layersConsists of a network of processing elements or nodes arranged in layers

Principle - Input pattern presented at the input layer causes network nodes to perform calculations in the successive layers until an output value is computed at each of the output nodes from which the most significant is selected

Working – Input to node j :Working – Input to node j : Output of node j :Output of node j : This continues through all the layers of the network until output layer is reached and output This continues through all the layers of the network until output layer is reached and output vector is computed vector is computed

The function f denotes activation function of each node. A sigmoid function is generally The function f denotes activation function of each node. A sigmoid function is generally used used

Extending the MLP methodologyExtending the MLP methodology Our neural network model used is a multilayer perceptron (MLP) network composed of one

input layer with four primary preprocessed variables (tPSA, fPSA, prostate volume, DRE) One hidden layer with two neurons, and one output layer with one neuron giving the output

value that is a measure of the probability of cancer.

13 parameters (weights) to be optimized. Activation function used in the hidden and output layers is hyperbolic tangent sigmoid function. Output values are between -1 to 1

[a= tanh(s) = (e^s – e^ -s) / (e^s + e^ -s)]

Formula for whole network : a1 = tanh (IW1,1*X1 + IW2,1*X2 + IW3,1*X3 + IW4,1*X4 + IB1)

a2 = tanh (IW1,2*X1 + IW2,2*X2 + IW3,2*X3 + IW4,2*X4 + IB2)

aout = tanh (LW1*a1 + LW2*a2 + LB)

After model build-up, initialize weights and enter inputs. Inputs are actual data of patients suffering from prostate cancer

Network now trained using Levenberg-Marquardt & Bayesian optimization techniques to recognize and associate patterns of input with desired outputs indicating the correct cancer risk.

Internal processing in network takes place and resultant output takes the form between 0 (low PCa risk) and 1 (high PCa risk). In some cases, the value is <0 or >1 which is not relevant

Intelligent learning is the key to ANN success. E.g. if the network has been trained for person having high PSA level due to non-cancerous cells then it wont indicate any cancer risk for persons falling into the same characteristic group

Training ProcedureTraining Procedure

5 Steps to Implement the ANN5 Steps to Implement the ANN

0

5

10

15

20

25

IstGroup

2ndGroup

3rdGroup

4thGroup

% fPSA

% tPSA

DRE

ProstateVolume

Step 1 – Obtaining the Input Data Set

Step 2 – Adjusting Initial Weights

i) model fitted on 75% and tested on 25% of patients in each training set.

ii) randomization, fitting and testing sessions repeated 5 – 6 times and the weights producing smallest sum of squared errors on the initial test set are selected as initial weights for the final training of the MLP.

Step 3 – Train Network

Step 4 – Test Network

Step 5 – Validating & Generalizing Network Efficiency

Actual Training of Network begins with estimation &

validation samples

If validation objective is acquired, network trained againtill best performance parameter values are obtained

Results of the ANN on a comparative basisResults of the ANN on a comparative basis

0102030405060708090

100

Sens Spec Acc

FPTPMLP

Sensitivity : Ability of the model to detect

prostate cancer early

Accuracy : Proportion of subjects with a correct

test result

Specificity : Efficiency of avoiding repeated

tissue samplings

ROC Curve : ANN model occupies greater area

under curve as compared to FP & TP models

Here comes our Final Product !Here comes our Final Product !

ConclusionConclusion Prostate cancer is the most common form of

cancer among men in the industrialized world

Screening is a very rough estimate for cancer risk

Further staging systems such as A,B,C,D, TNM lack required efficiency

Our model developed to aid or substitute the Our model developed to aid or substitute the current diagnosis and prognosis methodscurrent diagnosis and prognosis methods

Studies have shown that these highly accurate ANNs can Studies have shown that these highly accurate ANNs can detect prostate cancer early & reduce unnecessary tissue detect prostate cancer early & reduce unnecessary tissue samplingssamplings

Influencing factors :

a) Larger no. of input variables

b) Establishing interconnecting relationships b) Establishing interconnecting relationships

between thembetween them

c) Ability of network to be trained time and again c) Ability of network to be trained time and again

thereby increasing accuracy each timethereby increasing accuracy each time

ResultResult : Our proposed ANN model is far more efficient in : Our proposed ANN model is far more efficient in predicting prostate cancer risk and reduce the no. of false predicting prostate cancer risk and reduce the no. of false positive test resultspositive test results

© 2003 – The A.N.N.C.R.I.P.S

Join us to save a million lives

www.anncrips.com