ReTiNet: A cross-species few-shot learning framework for reticulocyte analysis under extreme data scarcity

Reticulocyte classification in peripheral blood smears is essential for assessing bone marrow activity, anemia progression, and chemotherapy response. Existing methods demand large labeled corpora and exploit cross-species weights purely as an initialization step, without any principled domain alignment mechanism. In this work, we propose ReTiNet, a unified cross-species meta-distillation framework that resolves these challenges within a single episodic learning architecture. A meta-trained ensemble of teacher networks learns transferable hematological priors from feline reticulocyte data, supervises a lightweight ShuffleNetV2 student via a three-component reticulocyte-focused distillation objective that jointly optimizes three-way classification, temperature-scaled KL-divergence knowledge transfer, and cross-domain feature alignment on reticulocyte queries through mean-squared-error embedding supervision. Quantization-aware training is embedded directly to learn efficiency constraints; post-training quantization is also evaluated to characterize the full efficiency-accuracy trade-off at deployment. The ShuffleNetV2 student distilled from a 3-teacher ensemble of MobileNetV2, EfficientNet-B0, and SqueezeNet1.1 achieves 99.52% reticulocyte-class accuracy and 97.0% F1-score without augmentation, and 93.13% overall accuracy with 93.17% F1-score under augmented conditions simulating real-world staining and imaging variability, surpassing the prior state-of-the-art ConcatNeXt while occupying just 1.46 MB, over 10 × smaller than comparable methods. Under extreme label scarcity, where only 5% of samples are labeled, ReTiNet achieves 83.41% accuracy and 74.19% F1-score without access to any large auxiliary corpus. INT8 quantization further compresses the student to 0.70 MB with minimal accuracy degradation. Model trustworthiness is validated through Grad-CAM++ and LIME, and further substantiated via cascading weight randomization analysis, confirming that attribution maps reflect learned morphological structure rather than spurious acquisition artifacts.