EAA-Net: Rethinking the Autoencoder Architecture with Intra-class Features for Medical Image Segmentation