Classificação hierárquica de imagens citológicas cervicais utilizando redes neurais convolucionais e ensemble learning

Abstract

Cervical cancer remains a significant challenge to global public health, particularly in low- and middle-income countries where access to screening programs is limited. Conventional diagnosis, based on the Pap smear, requires manual microscopic examination, a procedure that is labor-intensive and prone to errors. Although recent advances have expanded cytology automation, the combined application of hierarchical supervision and testtime augmentation (TTA) remains underexplored. This research proposes and evaluates a depth-aware hierarchical single-cell classification pipeline aimed at robustness, scalability, and semantic consistency. The study uses the HiCervix dataset, comprising 40,229 annotated cells distributed across 29 classes and three taxonomic levels, integrating convolutional and hybrid CNN–Transformer backbones, a hierarchical cross-entropy loss function, and frequency-adaptive data augmentation. At inference, the pipeline employs TTA with four visualization variants and probability fusion through ensemble to maximize predictive robustness. The results show consistent improvements, achieving accuracy of up to 78.9% at the species level and the lowest recorded hierarchical error (20.01%) among the trained models. Furthermore, the proposed low-complexity ensemble achieves accuracy comparable to state-ofthe-art models while requiring significantly lower computational cost.