I want to have a multilabel segmentation

[hannachoum]

HI there,

I have X-rays images that i need to segment.
I have 35 classes, but each pixel in the image can be part of several classes, so i need to make multilabel segmentation.
In the Annotation folder, i have npz object which is a concatenation of 35 matrix (composed with 0,1).
I need to modify the pipeline, in order to accept these objects.
Can you help me ?

Thank you for your help !

[HAOCHENYE]

Hi, sorry for my late reply. If you're using the mmsegmentation, I belive that you will get a faster and more accurate workaround there. If you're setting up your baseline directly based on MMEngine, the most simple way is to customize a Dataset that accepts your data format. The most important thing you need to do is to implement the __getitem__ method and return the expected output.

[hannachoum]

HI, thank you,
I already worked a lot on this case, but i need to rewrite all the transforms etc ... no ? Because i need them to accept my new data format.
do i also need to rewrite the loss of the model i'm using ?
I don't understand how changing only the getitem function can help me,
Thank you,

[HAOCHENYE]

Without knowing the specifics of your code implementation, it is hard to provide precise advice. However, if you are facing an issue where the input data format required by transforms does not match the raw data loaded by the dataset and the input expected by model.forward, the simplest solution is to convert the raw data to the format required by the transforms. Additionally, ensure that the function signature of model.forward aligns with the data returned by the transforms. For example, if the transforms return a dictionary with keys "inputs" and "data_samples," model.forward should accept these two parameters.