Medical images are available in small datasets with different modalities and for various organs. Although Transfer learning is a promising approach for training models on small datasets, further studies are required on using them with several image modalities and body parts. This paper explores two dominant meta-learning algorithms: a metric-based algorithm, namely Prototypical Network, and an optimization-based algorithm, namely MAML. The algorithms trained on a small multi-modal medical dataset (i.e., Slake), and a general dataset of the same size (i.e., Tiny-imagenet) with different class distribution methods: Random, Logical-based, and Statistical-based class distribution. The aim is to apply diversity among the classes in the meta-training set and similarity between the classes in the meta-training, and meta-testing or meta-validation sets. The results validate the importance of class distribution on the accuracy of the algorithms. MAML is hard to be trained on Tiny-imagenet but it shows good accuracy on Slake in specific cases. In statistical distribution, the Prototypical Network shows high accuracy on the datasets, especially when the similarity between the meta-training and meta-validation sets is considered for Tiny-imagenet.