Improving Robustness and Reliability in Medical Image Classification with Latent-Guided Diffusion and Nested-Ensembles