Bez tytulu slajdu

1
Rules for Melanoma Skin Cancer
Diagnosis Wlodzislaw Duch, K. Grabczewski, R.
Adamczak, K. Grudzinski, Department of Computer
Methods, Nicholas Copernicus University, Torun,
Poland. http//www.phys.uni.torun.pl/kmk
Zdzislaw Hippe Department of Computer Chemistry
and Physical Chemistry Rzeszów University of
Technology, zshippe_at_prz.rzeszow.pl
2
Content

• Melanoma skin cancer data
• 5 methods GTS, SSV, MLP2LN, SSV, SBL, and
  their results.
• Final comparison of results
• Conclusions future prospects

3
Skin cancer

• Most common skin cancer
• Basal cell carcinoma (rak podstawnokomórkowy)
• Squamous cell carcinoma (rak kolczystonablonkowy)

• Melanoma uncontrolled growth of melanocytes, the
  skin cells that produce the skin pigment melanin.
  
• Too much exposure to the sun, sunburn.
• Melanoma is 4 of skin cancers, most difficult to
  control, 179 Americans will develop melanoma.
• Almost 2000 percent increase since 1930.
• Survival now 84, early detection 95.

4
Melanoma skin cancer data summary

• Collected in the Outpatient Center of Dermatology
  in Rzeszów, Poland.
• Four types of Melanoma benign, blue, suspicious,
  or malignant.
• 250 cases, with almost equal class distribution.
• Each record in the database has 13 attributes.
• TDS (Total Dermatoscopy Score) - single index
• 26 new test cases.
• Goal understand the data, find simple
  description.

5
Melanoma AB attributes

• Asymmetry symmetric-spot, 1-axial asymmetry,
  and 2-axial asymmetry.
• Border irregularity The edges are ragged,
  notched, or blurred.Integer, from 0 to 8.

6
Melanoma CD attributes

• Color white, blue, black, red, light brown, and
  dark brown several colors are possible
  simultaneously.
• Diversity pigment globules, pigment dots,
  pigment network, branched strikes,
  structureless areas.

7
Melanoma TDS index

• Combine ABCD attributes to form one index
• TDS index ABCD formula
• TDS 1.3 Asymmetry 0.1 Border 0.5 S
  Colors 0.5 S Diversities
• Coefficients from statistical analysis.

8
Remarks on testing

• Test only 26 cases for 4 classes.
• Estimation of expected statistical accuracy on
  276 training test cases with 10-fold
  crossvalidation.Not done with most methods!
• Risk matrices desirable identification of Blue
  nevus instead Benign nevus carries no risk, but
  with malignant great risk.

9
Methods used GTS

• GTS covering algorithm (Hippe, 1997) recursive
  reduction of the number of decision rules.
• Interactive, user guides the development of the
  learning model.
• Selection of combination of attributes generating
  learning model is based on Frequency and Ranking.
• GTS allows to create many different sets of
  rules.
• In a complex situation may be rather difficult to
  use.

10
GTS results.

• GTS generated a large number (198) of rules.
• Experimentation allowed to find important
  attributes.
• Various sets of decision rules were generated
  TDS C-blue Asymmetry Border (4 attributes,
  based on the experience of medical doctors)TDS
  C-blue D-structureless-areas (3 attributes)
  TDS C-Blue (2 attributes)TDS (1 attribute) -
  poor results. Models with 2-4 attributes give
  81-85 accuracy.
• Combination and generalization of these rules
  allowed to select 4 simplified best rules.
• Overall 6 errors on training, 0 errors on test
  set.

11
Methods used SSV

• Decision tree (Grabczewski, Duch 1999)
• Based on a separability criterion max. index of
  separability for a given split value for
  continuous attribute or a subset of discrete
  values.
• Easily converted into a set of crisp logical
  rules.
• Pruning used to ensure the simplest set of rules
  that generalize well.
• Fully automatic, very efficient, crossvalidation
  tests provide estimation of statistical accuracy.

12
SSV results

• Pruning degree is the only user-defined
  parameter.
• Finds TDS, C-BLUE as most important.
• Rules are easy to understand IF TDS ? 4.85 ?
  C-BLUE is absent gt Benign-nevusIF TDS ? 4.85 ?
  C-BLUE is present gt Blue-nevusIF 4.85 lt TDS lt
  5.45 gt SuspiciousIF TDS ? 5.45 gt Malignant
• 98 accuracy on training, 100 test.
• 5 errors, vector pairs from C1/C2 have identical
  TDS C-BLUE.
• 10xCV on all data 97.50.3

13
Methods used MLP2LN

• Constructive constrained MLP algorithm, 0, 1
  weights at the end of training.
• MLP is converted into LN, network performing
  logical function (Duch, Adamczak, Grabczewski
  1996)
• Network function is written as a set of crisp
  logical rules.
• Automatic determination of crisp and fuzzy
  "soft-trapezoidal" membership functions.
• Tradeoff simplicity vs. accuracy explored.
• Tradeoff confidence vs. rejection rate explored.
  
• Almost fully automatic algorithm.

14
MLP2LN results

• Very similar rules as for the SSV found.
• Confusion matrix
• Original class Benign Blue- Malig-
  Suspi-
• Calculated nevus nevus nant
  cious
• Benign-nevus 62 5 0 0
• Blue-nevus 0 59 0 0
• Malignant 0 0 62 0
• Suspicious 0 0 0 62

15
Methods used FSM

• Feature-Space Mapping (Duch 1994)
• FSM estimates probability density of training
  data.
• Neuro-fuzzy system, based on separable transfer
  functions.
• Constructive learning algorithm with feature
  selection and network pruning.
• Each transfer function component is a
  context-dependent membership function.
• Crisp logic rules from rectangular functions.
• Trapezoidal, triangular, Gaussian f. for fuzzy
  logic rules.

16
FSM results

• Rectangular functions used for C-rules.
• 7 nodes (rules) created on average.
• 10xCV accuracy on training 95.51.0, test 100.
• Committee of 20 FSM networks 95.51.1, test
  92.6.
• F-rules, Gaussian membership functions 15 fuzzy
  rules, lower accuracy.
• Simplest solution should strongly be preferred.

17
Methods used SBL

• Similarity-Based-Methods many models based on
  evaluation of similarity.
• Similarity-Based-Learner (SBL) software
  implementation of SBM.
• Various extensions of the k-nearest neighbor
  algorithms.
• S-rules, more general than C-rules and F-rules.
• Small number of prototype cases used to explain
  the data class structure.

18
SBL results

• SBL optimized performing 10xCV on training set.
• Manhattan distance, feature selection TDS
  C_Blue
• 97.4 0.3 on training, 100 test.
• S-rules of the form IF (X sim Pi) THEN
  C(X)C(Pi)IF (TDS(X)-TDS(Pi)C_blue(X)-C_blue
  (Pi))ltT (Pi) THEN C(X)C(Pi) Prototype
  selection left 13 vectors (7 for Benign-nevus
  class, 2 for every other class.97.5 or 6 errors
  on training (237 vectors), 100 test
• 7 prototypes 91.4 training (243 vectors), 100
  test

19
Results - comparison
Method Rules Training Test SSV Tree,
crisp rules 4 97.50.3 100MLP2LN, crisp
rules 4 98.0 all 100 GTS - final
simplified 4 97.6 all 100 FSM, rectangular f.
7 95.51.0 1000.0 knn prototype
selection 13 97.50.0 100 FSM,
Gaussian f. 15 93.71.0 953.6 GTS initial
rules 198 85 all 84.6knn k1, Manh, 2 feat.
250 97.40.3 100LERS, weighted rules 21 --
96.2
20
Conclusions

• TDS - most important Color-blue second.
• Without TDS - many rules.
• Optimize TDS automatic aggregation of features,
  ex. 2-layered neural network.
• Very simple and reliable rules have been found.
• S-rules are being improved - prototypes obtained
  from learning instead of selection.
• Data base is expanding need for non-cancer data.