You are on page 1of 24

Artificial Neural Networks In Medical Diagnosis

Definition
Artificial Neural network is an information processing system composed of highly interconnected processing elements (neurons) working in union to solve specific problems.

The general structure of artificial neural networks consists of neurons interconnected to form three types of layers:
Input Layer (Receives input from outer world) One or more Hidden Layers (Process input) Output Layer (Gives output to outer world)

Biological Inspiration
ANN is Highly interconnected set of neurons. Inspired from biological neurons. Output is produced by applying activation function on the input.

Activation functions
20 18 16 14 12 10 8 6 4 2 0

Linear Function Linear (x) = x

10

12

14

16

18

20

2 1.5 1 0.5 0 -0.5 -1 -1.5 -2 -10

Logistic Function Logistic (x) = 1/(1+exp(-x))

-8

-6

-4

-2

10

Sign Function Sign (x) = +1 if x >= 0, else -1

Step Function Step (x) = 1 if x >= t, else 0

Transformation
How input is transformed to outputs:

Architecture
Neurons are linked together in layers to form a specific architecture. Transforms inputs into meaningful outputs. Artificial Neural network can be broadly of two types:
Feed Forward Architecture Recurrent Architecture

Feed Forward Neural Networks


The information is propagated from the inputs to the outputs Computations of Non linear functions from n input variables. There is NO cycle between outputs and inputs.
Output layer

2nd hidden layer 1st hidden layer

x1

x2

..

xn

Recurrent Neural Networks


0 0

1 0
1 0

Can have arbitrary topologies. Training is more difficult. There can be cycles between outputs and inputs.

x1

x2

Training
Every neuron in ANN has a weight associated with it. Training takes place by giving different instance inputs. Weights are adjusted until the desired results are got as output.

Training can be performed in two modes:


Supervised Unsupervised

Some training algorithms of common use are: BPN, MLP, RBF, LVQ etc. Training is continued until a desired accuracy is achieved or error comes to a minimum level.

Neural Networks in Medical Field


Disease Diagnosis Prediction of chances of a disease Robotic Handling of Surgery Measuring Effect of new Medicines Continuous Patient Monitoring

Collect Samples
Take both normal and abnormal samples of images or data from tests. More the number of samples, more will be accuracy. Type of image depends upon the nature of research.

Preprocessing
Preprocessing is necessary for rendering the image fit for analysis. If one is using data then image processing steps are not required. Various methods used by different researchers are:
Enhancement Filtering Compression etc..

Extract Features
Extract features of the image that are necessary for diagnosis. Difference between left and right eye features will be used for diagnosis. In case of data, important parameters are selected.

Normalize
A data set is created by normalizing the features and actual diagnosis results.
S. No. Features Result

Training And Testing Set


Dataset can be randomly divided into mutually exclusive training and testing sets. Training set is usually larger than the testing set. Size depends on number of available samples.

K-Fold Cross Validation


Another method for selecting training and testing sets. Samples are divided into K mutually exclusive subsets of equal size. K-1 subsets are used for training and one subset is used for testing. The procedure is repeated K times taking different subset for testing.

Design of Neural Network


Number of inputs and outputs. Number of hidden layers. Number of neurons in each hidden layer. Interconnections between different layers. Training algorithm to be used. Initial weights.

Training and Testing


Input the training set along with the results. Training is performed until weights continue to change. After training, test set is introduced but the results are not fed along with. Now the neural network has to perform similar operations and produce diagnosis.

System Evaluation
Methods such as confusion matrix can be used to evaluate system performance.

Evaluation Parameters
Sensitivity=

Specificity=

Accuracy=

THANKS

You might also like