Residual sequential attention and class-imbalance learning for multi-class peripheral blood cell classification

Abstract Peripheral blood cell classification is challenging when severe class imbalance coincides with fine visual differences among maturation stages and morphologically related cell types. This study investigates a residual sequential attention framework built on ResNeSt50 and inverse-frequency weighted categorical cross-entropy for 13-class classification on the KU-Optofil PBC dataset. The attention block retains the standard CBAM channel and spatial operators, applies them sequentially after the final residual stage, and fuses the refined representation with the incoming feature map through an additive residual connection. No morphology supervision or localization annotation is used, so the attention branches are interpreted as feature-refinement operations rather than direct identifiers of cytoplasmic or nuclear structures. Under the common from-scratch protocol, the final configuration reaches 99.43% accuracy with a 95% confidence interval from 99.21 to 99.59, 99.44% weighted precision, 99.43% weighted recall, 99.43% weighted F1-score, 96.81% macro F1-score, and 97.09% balanced accuracy. Relative to plain ResNeSt50, the observed single-run improvements are 3.37 percentage points in accuracy and 3.48 points in weighted F1-score. Class-wise confidence intervals show substantially greater uncertainty for the rarest categories, indicating that aggregate accuracy alone is insufficient for judging performance under the long-tailed distribution. The results support strong performance within the released test split, while the single-run design and absence of independent external evaluation limit broader generalization.