Accurate classification of hematological cell types in microscopic images is essential for reliable diagnosis and screening; however, performance in real-world settings is frequently limited by annotation scarcity, class imbalance, and subtle inter-class similarity. To mitigate these constraints, we propose a self-supervised pretraining (SSP) Vision Transformer framework based on an iBOT teacher–student distillation strategy, designed to learn transferable representations from unlabeled microscopy images prior to task-specific supervision. The method optimizes complementary objectives at two granularities: (i) global semantic learning via the class token and (ii) fine-grained morphological learning via masked patch-token distillation, thereby strengthening feature robustness for downstream classification. The pretrained backbone is then fine-tuned with a lightweight classification head and evaluated under a controlled data-efficiency protocol using an eight-class hematology microscopy dataset comprising 17,092 images, with a fixed held-out test split of 3,421 images and progressively reduced labeled training partitions (100%, 90%, 80%, 70%, 60%, and 50% of the remaining data). Across multiple backbone scales (ViT-S/16, ViT-B/16, and ViT-L/16), SSP pretraining consistently improves performance over purely supervised Vision Transformer baselines in accuracy, precision, recall, specificity, F1-score, with particularly strong gains for minority and visually ambiguous classes, and with advantages that become more pronounced as labeled data decreases. Importantly, these improvements are achieved without increasing model capacity, since parameter counts remain identical to their supervised counterparts, at the cost of a moderate and explicitly reported pretraining overhead. Overall, the proposed framework provides a scalable, annotation-efficient transformer-based solution for microscopy classification in resource-constrained biomedical environments.
Rahman et al. (Sun,) studied this question.