A%20Computer-Aided%20Diagnosis%20System%20For%20Digital%20Mammograms%20Based%20on%20Radial%20Basis%20Functions%20and%20Feature%20Extraction%20Techniques

1
A Computer-Aided Diagnosis System For Digital
Mammograms Based on Radial Basis Functions and
Feature Extraction Techniques

• Dissertation written by
• Mohammed Jirari
• Proposal for Ph.D. Candidate Examination
• July 23rd, 2003

2
Why This Project?

• Breast Cancer is the most common cancer and is
  the second leading cause of cancer deaths
• Mammographic screening reduces the mortality of
  breast cancer
• But, mammography has low positive predictive
  value PPV (only 35 have malignancies)
• Goal of Computer Aided Diagnosis CAD is to
  provide a second reading, hence reducing the
  false positive rate

3
What is a Mammogram?

• A Mammogram is an x-ray image of the breast.
  Mammography is the procedure used to generate a
  mammogram
• The equipment used to obtain a mammogram,
  however, is very different from that used to
  perform an x-ray of chest or bones. The breast is
  composed of tissues that are similar to each
  other in density. Changes or abnormalities in the
  breast tissue are often very subtle. Therefore,
  the mammogram machines, film, and developing
  process are specially designed to take pictures
  of these subtle differences.

4
Mammograms (cont.)

• In order to get a good image, the breast must
  also be flattened or compressed. This may be
  uncomfortable, but it will not harm the breast in
  any way and is extremely important for obtaining
  a clear image. Compression of the breast is also
  beneficial because it results in a lower dose of
  radiation.
• In a standard examination, two images of each
  breast are taken--one from the top (called a
  cranio-caudal or CC view) and one from the side
  (called a medio-lateral oblique or MLO view).
  This ensures that the images display as much
  breast tissue as possible.

5
Mammogram Examples

Mammogram of a left breast, cranio-caudal (from
the top) view
Mammogram of a left breast, medio-lateral oblique
(from the side) view
6
Purpose of CAD

• Mammography is the most reliable method in early
  detection of breast cancer.
• But, due to the high number of mammograms to be
  read, the accuracy rate tends to decrease.
• Double reading of mammograms has been proven to
  increase the accuracy, but at high cost.
• CAD can assist the medical staff to achieve high
  efficiency and effectiveness.
• The physician/radiologist makes the call not CAD

7
Proposed Method

• The proposed method will assist the physician by
  providing a second opinion on reading the
  mammogram, by pointing out an area (if one
  exists) delimited by its center coordinates and
  its radius.
• If the two readings are similar, no more work is
  to be done.
• If they are different, the radiologist will look
  at it one more time to make the final diagnosis.

8
Co-occurrence Matrices

• The joint probability of occurrence of gray level
  a and b for two pixels with a defined spatial
  relationship in an image.
• The spatial relationship is defined in terms of
  distance d and angle ?.
• From these matrices, a variety of features may be
  extracted.

9
Co-occurrence Matrices (cont.)

• In my project, the matrices are constructed at
  distance of d1 and d3 and for angles ?0, 45,
  90, 135.
• For each matrix, eight features are extracted.
• Can be formally represented as follows

10
Features Used

• Energy or angular second moment
• Entropy
• Maximum Probability
• Inverse Difference moment
• ?2, ?1

11
Features Used (cont.)

• Contrast
• Homogeneity
• Inertia or variance

12
Features Used (cont.)

• Correlation

13
Radial Basis Network Used

• Radial basis networks may require more neurons
  than standard feed-forward backpropagation FFBP
  networks
• BUT, can be designed in a fraction of the time to
  train FFBP
• Work best with many training vectors

14
Radial Basis Network with R Inputs
15
aradbas(n)
16
Radial basis network consists of 2 layers a
hidden radial basis layer of S1 neurons and an
output linear layer of S2 neurons
17
Data Used in my Project

• The dataset used is the Mammographic Image
  Analysis Society (MIAS) MINIMIAS database
  containing Medio-Lateral Oblique (MLO) views for
  each breast for 161 patients for a total of 322
  images.
• Every image is
• 1024 pixels X 1024 pixels X 256

18
Preprocessing

• In order to improve the quality of the images and
  make feature extraction more reliable, 2
  techniques were used
• Cropping cuts the black parts of the image
  (almost 50) based on a threshold
• Enhancement Histogram equalization to
  accentuate the features to be extracted by
  increasing the dynamic range of gray levels

19
Preprocessing result
a-Original mammogram b-after cropping c-after
cropping and histogram equalization
20
Feature extraction

• The extraction phase is needed in order not to
  feed the whole image as input to the neural
  network. The method applied takes the whole
  cropped image and calculates the co-occurrence
  matrices at distance d1 and d3. The angles
  used are ?0, 45, 90, 135 with the fifth
  matrix being the mean of the 4 directions. The
  co-occurrence matrices are calculated and the
  eight statistical features mentioned earlier are
  computed.

21
Training

• After normalizing the data between 0 and 1 for
  the network to have a common range, the training
  begins.
• The first training set was made up of 212
  mammograms with 81 abnormal ones, with features
  calculated at distances d1 and d3.
• The second training set was made up of 163
  mammograms with 81 abnormal ones, with features
  calculated at distances d1 and d3.

22
Example of a network used
23
Testing

• A mammogram is presented to the trained network
  and the output is a suspicious area denoted by
  its centers x and y coordinates and its radius.
  If the mammogram is considered to be normal then
  zeros are returned for the coordinates and
  radius.
• The radiologist can then review his original
  assessment of the patient if some areas uncovered
  by the network were not originally looked at
  closely.
• The whole database is tested and the accuracy is
  calculated.
• The smaller dataset performed better than the
  larger, and when d3 results were better also
  compared to d1.

24
Results

• There were 2 training datasets 163 and 212
• There were 2 distance measures 1 and 3
• There were 3 spreads 0.1, 0.25, and 0.05
• There were 3 goals 0.00003, 0.008, 0.00005.
• For 12 possible combinations.
• The NN was sensitive to the unbalanced data
  collection that contained about 70-30 split in
  the larger training set. Therefore the smaller
  dataset was preferred.
• Achieving a high recognition is not that
  appealing if the TPF is small

25
Representative Preliminary Results
Net 1 Net 2 Net 3
TPF 0.01639 0.72973 0.88043
FPF 0.5939 0.0 0.3478
Recognition 0.3323 0.9068 0.7174
of Neurons 133 91 151
26
Future work/plans

• Use more features like standard deviation,
  skewness, kurtosis, ...
• Which feature(s) have the most impact
• Rank the features from best to worst (single
  input to NN)
• Select most significant feature(s) by using
  leave one out method
• Determine whether the area is benign or malignant
  by adding the severity of the abnormality to the
  training.

27
Future work/plans (cont.)

• Try and reduce False Negatives on the basis of
  region characteristics size, difference in
  homogeneity and entropy.
• Use larger database to train/learn, since most
  commercial CADs use 100,000s mammograms to try
  and recognize foreign samples .
• Increase the recognition rate to diagnose with
  100 accuracy since saving human lives is at
  stake. Reaching 80 rate determines credibility
  of CAD. May or may not be reached when tested on
  foreign mammograms, but can gain valuable ideas
  as to how to improve.

28
Future work/plans (cont.)

• Use segmentation of breast from its background as
  it may make the feature extraction more accurate.
• May experiment with multichannel wavelet
  transform and Kalman-filtering NN, since wavelet
  transform provides an efficient multiresolution
  representation. I may also experiment with a
  fuzzy neural CAD using a fuzzy detection
  algorithm using a sliding window. Comparing the
  results will be worth investigating.
• Will use some data mining techniques to unveil
  new patterns/relationships between the presented
  patients.