Dual-Domain Calibration with Source Rehearsal and Knowledge Distillation for Robust Cross-Dataset Classification

Abstract

Automated peripheral blood cell classification has shown strong performance on curated microscopic image datasets, but model reliability can decrease when images are acquired from external sources with different visual characteristics. This study presents a reproducible deep learning protocol for five-class peripheral blood cell morphology classification involving basophils, eosinophils, lymphocytes, monocytes and neutrophils. The protocol compared five pretrained architectures, including Swin-Tiny, ConvNeXt-Tiny, MaxViT-Tiny, DeiT-Small and EfficientNetV2-B0, using balanced PBC training, validation and locked-test splits, followed by external zero-shot testing, image-quality stress evaluation and external-domain calibration. Initial PBC-only training produced high internal test performance but substantially lower external zero-shot performance, indicating a dataset-shift problem. To address this, we developed a source-retaining dual-domain calibration strategy that combines external calibration data, PBC rehearsal samples and teacher distillation. DeiT-Small was selected as the calibration backbone because it provided the strongest balance between PBC performance, stress robustness and external zero-shot generalisation. On locked testing, the proposed approach achieved a macro-F1 score of 0.9933 on the PBC test set and 0.9746 on the external test set. External class-wise performance remained high across all five cell categories, with macro precision, recall and F1 of approximately 0.9747, 0.9747 and 0.9746, respectively. These findings suggest that source-retaining calibration can improve cross-dataset generalisation while preserving source-domain performance. The study provides a transparent evaluation framework for assessing robustness, external validity and domain-shift effects in microscopic blood-cell image classification.