Latent Representation and Generative Augmentation with Graph-Based Learning for Imbalanced Leukocyte Cytomorphology Classification

In this work, we propose a hybrid framework that integrates autoencoder-based latent representation, GAN-driven minority-class augmentation, and graph-based classification with GraphSAGE. Our key contributions are as follows:
1. We propose a latent-space augmentation strategy that synthesizes minority-class embeddings within the autoencoder feature manifold, avoiding artifacts common in pixel-level augmentation.
2. We design a graph-based relational learning framework that embeds leukocyte representations into a similarity graph and applies inductive classification to leverage contextual neighborhood information.
3. Through extensive experiments on the AML-Cytomorphology-LMU dataset, our method achieves ~91% accuracy and improved macro-F1 scores, particularly enhancing recall for rare subtypes.
4. We show that latent augmentation improves minority-class sensitivity and reduces cross-validation variability, yielding more robust and clinically deployable models.