Fix class imbalance errors by loss function weighting

[Sajtospoga01]

Try calculating relevant loss to improve prediction performance on minimal dataset

• inverse logarithmic based on overall class frequency
• tf-idf vector like dynamic weighting function
• loss function weighting in hybrid loss (e.g.: structural similarity is down weighted if there is one structure dominating the picture)