Medical applications of learning machines offer a challenge for both medical doctors and mathematicians/engineers. The main question arises when you decide to teach the model the same way a human being understands slight similarities and differences or improve the human (medical) capabilities for interpretation. Mathematically speaking, this translates on the artifact removal, preprocessing and contrast for the spatial frequencies on the images and their relationship with the complexity of the learning machine looking for the generalization and robustness of the solution for proper artificial intelligence based decision making. This paper addresses these challenges and provides insights on their solution in real life scenario. Particularly, we offer a discussion on the knowledge database built from medical expertise and take care of the mathematical boundaries to be satisfied by the models. Also, we consider during our approach several restrictions from the authorities for the medical tool deployment using them as part of the indexes of performance.