change 'classWeights' in unetLayers

Hello,

im using Unet to perform semantic segmentation (I want to segment liver metastasis). I want to weight 'more' the pixels of the metastasis respect to the background but I can't manage to do it...

Code for creating unet architecture:

lgraph = unetLayers(imageSize, numClasses, 'EncoderDepth', 4);

How can I change the 'classWeights' in lgraph.Layers(end).pixelClassificationLayer in order to do this??

If, for example, I do this:

lgraph.Layers(end).pixelClassificationLayer.classWeights = [2, 0.5];

I receive an error telling me that i have first to put the 'labels' to the 'classes' options in lgraph.Layers(end), but if i do this:

lgraph.Layers(end) = {'metastasis', 'background'};

I receive an error telling that that layer is in 'read only' mode....

Please im getting crazy...

Thanks in advance for any help!

Answer 1

[Raunak Gupta]

Hi,

The unetLayers object in MATLAB is in read only mode and we cannot change the properties of layers as such. You may remove the last layer which is the Segmentation-Layer and can add a new Segmentation-Layer with the properties that are mentioned. Following Code will help.

% Last Layer Name is 'Segmentation-Layer' 
% for checking layer name use :- lgraph.Layers 
lgraph = removeLayers(lgraph,'Segmentation-Layer'); 
layerlast = pixelClassificationLayer('Classes',{'metastatis','background'},'ClassWeights',[2,0.5],'Name','New_segmentation_Layer'); 
layer_to_add = [layerlast]; 
lgraph = addLayers(lgraph,layer_to_add); 

Reference to the functions that are used removeLayers, pixelClassificationLayer, addLayers.

[Marco Festugato]

It worked! Thank u so much :)

[Timothy Lee]

By using pixelClassificationLayer with class weights, is this the same as implementing a custom weighted classification layer, as shown in this example:

https://de.mathworks.com/help/deeplearning/ug/create-custom-weighted-cross-entropy-classification-layer.html