Dictionary learning for extremely small dataset of samples?

I have around 100 samples, on which I'd like to do some sort of "dictionary learning". These are essentially medical images of a particular ethnicity and I cant really augment the data. I am looking to try a dictionary learning approach where I can use the learnt dictionaries on "healthy specimens" to reconstruct test data and discover abnormalities.

I am wrangling with the question of which method could make the best use of this small amount of data when learning to do reconstruction? Normally for classification tasks people take the "tendency towards overfitting" as a deciding factor between schemes. Is there an analogue for reconstruction type problems?

Apart from getting our hands dirty and trying out many different methods (sparse coding, NMF, autoencoders etc.) which I expect I will have to, what would be a good direction to think about this?

thanks guys.