Transfer Learning is currently popular in Medical Image classification. Transfer Learning methods are extensively applied with CNN’s such as Res-net, Densenet, VGG16, Inception, etc. for various medical diagnosis tasks. CNN’s are around since the 1980s, but 60-80 percent of the TL research in MIC is done in the last three years. While CNN’s can be traditionally used as they are, they have been ensembled, segmented and improvised in recent days to resolve multiple MIC problems. This Review identified three main challenges in implementing Transfer Learning for Medical Image Classification (1) Overparameterization of deep CNN’s (2) Expensive Computations and (3) Insufficient availability of labeled data in the Medical field. The study also identified the opportunities in the form of Light-weight architectures and Multi-stage Transfer Learning which could potentially mitigate the above-mentioned challenges.