How to improve my GLCM results to get best result using ANN

Hi everyone

I am studying on mammogram images . I am using glcm for feature extraction .

I have two codes , one of gives 88 features another is 7.

I am using that 7 different number for as an input and I have 1 column and writing three different number in that column for my outputs.

normal situation is 0,

benign situation is1,

malign situation is 2,

according to that 7 numbers as an input and for 3 different number for outputs (8th column for output)

my ANN (articial neural network) result is very bad.

How can I improve my ANN result ?

When I create inputs regularly myself , my ANN result is perfect , but according to GLCM numbers is bad.

Something wrong should be according to glcm results but ı couldn't find.

What can I do ? Any suggestions?

I added my datas excel file. First 7 column is input, 8th column is output .( first row has 7 inputs and its equavalnce is 8th column.)

Answer 1

[Madhav Thakker]

Hi Ali,

I understand you are getting bad results with ANN. You might want to look into the 7 feature values you obtain using glcm. Are these values capable of differentiating the classes?

That being said, one possible reason for that could be the dataset size. There are only 31 values in the file attached. This might not be sufficient to train a good network. You can try to accumulate more data, or using a smaller network.

Another reason could be class imbalance. There are only 3 samples with class label 2. You can look into https://in.mathworks.com/help/stats/classification-with-imbalanced-data.html for classification with imbalanced data.

Hope this helps.

[Ali Zulfikaroglu]

Yeah I know data is not enough , I tried with more data but the result was same. I was using neural net fitting application on matlab and I also have codes about ANN. But as I mentiomed the results was bad. And then I applied neural net pattern recognition application and result was pretty good. But row and column should be shift(transpoze actually) for that application . I think I solved my question but thank you for answering and ideas.