NEURAL NETWORK BASED SYSTEM FOR

1

• NEURAL NETWORK BASED SYSTEM FOR
• DECISION MAKING SUPPORT
• IN ORTHODONTIC EXTRACTIONS

R. Martina1, R. Teti2, D. DAddona2, G.
Iodice1 1Dept. of Odontostomatology, University
of Naples Federico II, Naples, Italy 2Dept. of
Materials and Production Engineering, University
of Naples Federico II, Italy
2
Introduction

• Dental malocclusions are a highly prevalent
  pathology in the general population and the
  greater attention to aesthetic and functional
  problems have driven to a larger demand of
  orthodontic treatment in the last years
• A critical step in the orthodontic therapy is
  represented by a correct diagnosis and treatment
  planning
• Orthodontic diagnosis, however, often results
  very difficult and influenced by subjective
  interpretation of the measured parameters
• For this reason, Artificial Neural Network (NN)
  based approaches have been proposed as a valid
  support for diagnosis in orthodontics
• Treatment planning is a decisive and critical
  moment for the clinician, especially in case of
  extraction, which is a non reversible procedure

3
Introduction

• The diagnostic parameters needed to define the
  treatment plan are very numerous and the relative
  weight of each parameter for a specific patient
  is not easy to determine
• This subjective aspect of orthodontic diagnosis
  determines the lack of universality and unanimity
  in the interpretation of orthodontic data and in
  the treatment selection
• A referencing tool for orthodontic data
  evaluation would be desirable, particularly in
  controversial cases, where subjective data
  interpretation could generate incorrect decisions
• An innovative and promising approach to this kind
  of problems is represented by the development and
  application of intelligent computation procedures
• The main objective of this work is the
  realization of an intelligent diagnostic system
  based on NN for the extractive therapeutical
  option on the basis of measured orthodontic
  parameters.

4
 
Artificial Neural Network (NN) Architecture

• An artificial Neural Network (NN) is a
  computational model of the human brain that
  assumes that computation is distributed over
  several simple interconnected processing
  elements, called neurons or nodes, which operate
  in parallel
• Feed-forward back-propagation NN (BP NN)

5
 
Artificial NN Functioning

• The outputs of nodes in one layer are transmitted
  to nodes in another layer through connections
  that amplify or attenuate the outputs through
  weight factors.
• Except for input layer nodes, the net input to
  each node is the sum of the weighted outputs of
  the nodes in the prior layer.
• Each node is activated in accordance with the
  input to the node, the activation function of the
  node, and the threshold of the node accordingly,
  the node fires an output
• A NN provides a mapping through which points in
  the input space are associated with corresponding
  points in an output space on the basis of
  designated attribute values, of which class
  membership might be one

6
 
Artificial NN Functioning

• NNs can capture domain knowledge from examples,
  do not archive knowledge in an explicit form such
  as rules or databases, can readily handle both
  continuous and discrete data, and have a good
  generalisation capability
• Knowledge is built into a NN by training. Some
  NNs can be trained by feeding them with typical
  input patterns and expected output patterns
• The error between actual and expected outputs is
  used to modify the weight of the connections
  between neurons. This method is called supervised
  training

7
 
NN Data Processing

• A supervised NN paradigm based on a feed-forward
  back-propagation (BP) NN is employed to provide
  for the clinical decision on orthodontic
  extractions
• The training set utilised for NN learning was
  built on the basis of orthodontic casts and
  radiographic measurements as well as clinical
  examinations carried out on 48 patients
• Cephalometric analysis was performed on lateral
  standardized cephalograms, taken by a single
  technician using the same x-ray device and a
  standardized procedure
• The cephalograms were made with the mandible in
  the intercuspal position.

8
 
Orthodontic Casts and Radiographic Measurements

• Cephalometric analysis was performed on lateral
  standardized cephalograms, taken by a single
  technician using the same x-ray device and a
  standardized procedure
• The cephalograms were made with the mandible in
  the intercuspal position

9
 
NN Data Input

• Orthodontic casts and radiographic measurements
  32 features

10
 
NN Configurations

• The 32 features made up a 32-component input
  vector and the extraction therapeutical option
  represented the corresponding 1-component output
  vector classified as
• extraction (Od 1)
• not extraction (Od 0)
• The input and output features made up a
  33-component pattern vector for each patient or
  case
• Several NN configurations were tested
• Results are presented for the 32-4 or 8 or 12 or
  16-1 NN configurations
• - input layer with 32 nodes for the 32-component
  input vector
• - hidden layer with 4-8-12-16 nodes
• - output layer with 1 node for extraction option

11
 
NN Parameters

• The 32-4 or 8 or 12 or 16-1 NN main parameters
  were
• - weights and thresholds randomly initialized
  between -1 and 1
• - learning coefficient ?? 0.3 or 0.15
• - momentum a 0.4
• - learning rules Normal Cumulative Delta Rule
  and Cumulative Delta Rule
• - transfer functions sigmoid function and
  hyperbolic tangent function
• - learning steps for a complete training 10,000
  - 100,000
• - epoch size was 6.

12
 
NN Processing

• NN training and testing was performed by the
  leave-k-out method one case was kept aside in
  turn while the remaining were used for training.
  The kept aside case was then used for testing
• This procedure was repeated for all cases in the
  training set
• In the training phase the 32-component input
  vectors were presented at the input layer of the
  32-4 or 8 or 12 or 16-1 NN and the extraction
  option was fed into the output layer
• In the testing phase the 32-component input
  vectors were presented at the input layer of the
  32-4 or 8 or 12 or 16-1 NN and the extraction
  option was expected at the output layer

13
 
NN Results

• The NN output was considered correct if its
  output (extraction or not extraction) coincided
  with the decision made for the patient at the
  moment of treatment
• The NN output was correct if the actual output
  was lower than 0.5 for a not extraction case or
  higher than 0.5 for an extraction case
• The NN error was calculated as Er (Oa - Od)/1,
  where
• - Oa actual output,
• - Od desired output
• 1 difference between two adjacent numerical
  code values
• In case of a not extraction case, if the output
  was lower that -0.5 it was considered equal to
  -0.5 and, on the contrary, in the case of an
  extraction case, if the output was over 1.5 it
  was considered equal to 1.5
• The expected extraction diagnosis was correct if
  Er was between -0.5 and 0.5 otherwise, a
  misclassification case occurred

14
 
Neural Network Processing and Results 32-4-1

• The first NN configuration had structure 32-4-1
• - input layer with 32 nodes for 32 patient
  features
• - hidden layer with 4 nodes
• - output layer with 1 node for the extraction
  option

15
 
Neural Network Processing and Results 32-8-1

• The first NN configuration had structure 32-8-1
• - input layer with 32 nodes for 32 patient
  features
• - hidden layer with 8 nodes
• - output layer with 1 node for the extraction
  option

NN Output (? Not extraction ? Extraction)
NN Error (? correct classification case,
misclassification case)
16
 
Neural Network Processing and Results 32-12-1

• The first NN configuration had structure 32-12-1
• - input layer with 32 nodes for 32 patient
  features
• - hidden layer with 12 nodes
• - output layer with 1 node for the extraction
  option

NN Output (? Not extraction ? Extraction)
NN Error (? correct classification case,
misclassification case)
17
 
Neural Network Processing and Results 32-16-1

• The first NN configuration had structure 32-16-1
• - input layer with 32 nodes for 32 patient
  features
• - hidden layer with 16 nodes
• - output layer with 1 node for the extraction
  option

NN Output (? Not extraction ? Extraction)
NN Error (? correct classification case,
misclassification case)
18
 
Neural Network Success Rate

• The NN performance (success rate) was calculated
  as the ratio of correct identifications over the
  total pattern vectors presented at the NN input
  for all NN configurations considered 32-4 or 8
  or 12 or 16-1

NN success rate for different number of nodes in
the hidden layer
NN success rate vs. number of nodes in the hidden
layer
19
 
Conclusions

• This work shows that neural network based
  decision making support systems may be trained on
  the basis of clinical data
• The decision making support systems can be used
  where rule based decision making is unreliable
  or impossible this is the case in many clinical
  situations, especially in orthodontic treatment
  planning
• Neural network systems may, therefore, become an
  important decision making support tool in
  orthodontics and find applications both in
  improving the clinical treatment and in
  maximizing the cost benefit of the treatment