I have an n-takes k-shots medical image segmentation problem.
-Tasks: Different human organs ex: liver, spleen, kindness etc...
-Shots: 10 CT scans NIFTI images, where all tasks(human organs) exist in all shots, but one of them is labelled to match the task.
-Meta-teasing and meta-training have only one human organ segmentation according to the task. For example, Task 1 is learning the liver only since the segmentation is just the liver. Task 2 is learning the spleen only since the segmentation is just the spleen.
-Final theta is tested using n images. Each image has the segmentation of all tasks.
Now my question is, what is the right way to do in this case? Because when the model could generalise better, the loss function will be worse, and the model will not be good. In other words, the updated theta should generalise for all human organises. In meta-training and meta-testing, the theta will focus on the human organ already segmented for that task and assume the other organs already existing in the images are false.
I am unsure if I am true, but I understand this from theory and implementation. Correct me if I am wrong, and please explain.
What I am training to do to avoid the problem:
1- Crop the images to cover only one task ( It may work for limited and not overlapped tasks). I can't apply it since I have overlapped human organs.
2- Use a loss function for meta-training and meta-testing to measure only the true positive and false negative.
I am thinking loud and writing a lot. Sorry about that, but I got lost, and I need your opinion.
