Which's the better classifier to use for skin issues identification using matlab

My project consists of extracting color,texture and shape features and try to identify skin conditions,i'll obviously need more than 10 classes(diseases) and have around 40 to 50 samples of each to train,i've done my research regarding Neural networks,KNN,Naive Bayes,Multi class SVM and others but i'm still indecisive regarding which to use knowing that the most previous researches in this field are usually 2 classes using SVM or knn,so any advice or guidance would be appreciated.

Answer 1

[Greg Heath]

I don't know which would be best for you. However, NNs are universal approximators and I would be surprised if any of the others would have significantly lower error rates.

The appropriate NN to use is PATTERNNET. See the online documentation

 help patternnet

and

 doc patternnet

I have posted many examples on both the NEWSGROUP and ANSWERS. Use the search words

 greg patternnet

Hope this helps.

Thank you for formally accepting my answer

Greg

[Elias Unk]

Hi greg and i've checked the keywords you provided,it's defo interesting however i didn't find much examples regarding NN training to detect skin based issues nor reasons/studies about why it's any better or superior than other classifiers in that exact field,the best i could find using the NN is a 2 class study to detect a plant's issue from it's leaf.A solid matlab based database example or a recent study link where NN was used to classify skin lesions shall be appreciated if you have any.

Answer 2

[Greg Heath]

PATTERNNET is a universal classifier which typically means for a given set of reasonable input/target pairs you will probably do no better with another type of classifier.

On the other hand, another type of universal classifier like NEWRB might be easier to train and understand.

Hope this helps.

Greg

[Elias Unk]

It has,i think i'll end up using it and other classifiers and compare my results as Image Analyst suggested.

Answer 3

[Image Analyst]

Try using perfcurve() in the Statistics and Machine Learning Toolbox to compute the ROC curve for the various techniques.

You can also use confusionmat() to see how well your 10 classes are being properly classified. Ideally you'd have a diagonal matrix. The wider it is, the less accurate the technique is.