I am trying to implement a segmentation model for medical images. my question is should I train the segmentation model with both the normal and up-normal samples. Or each one in a different model(the normal and abnormal samples)? I am a little confused about this.

CodePudding user response：

Its a decision about how you want your software to work. You can train both samples together if situation requires it and maybe lower accuracy is acceptable. If you dont mind a slower software, you can stick a classifier on the pipeline to decide if image is normal or abnormal and pass through segmentation model it matches with. Its a matter of your resources and requirements.