Federated Model Aggregation via Self-Supervised Priors for Highly Imbalanced Medical Image Classification