I'm developing lung segmentation program in MATLAB, not being allowed to use any sort of machine learning, and ideally implementing all (or most) image processing methods I use, for an assignment.

Currently, it's mostly based on Otsu thresholding and morph. transformations and (kinda) good results look something like this: Example 1 - Original, Example 1 - Segmented, Example 2 - Original and Example 2 - Segmented.

Not good cases look like this: Example 3 - Segmented and actual failures look like this: Example 4 - After Otsu.

What king of processing (including pre or post) could I use to improve the lungs' outlines in good cases?

How can I correcly fill the mask in the not good scenario? (I've messed around with imfill and it didn't help much)

And, at last, is there a way to ensure that the lungs will always be a closed region after Otsu thresholding? (Failure case).

Any help would be greatly appreciated!