Attention-integrated hybrid CNN–vision transformer framework for robust multi-class cervical cytology classification using liquid-based cytology images

The grading of cervical cytology images is a challenging task because some morphological changes are subtle, there are class imbalances and variability in staining of lesion grades. In order to overcome such difficulties, this work suggests MAH-CerviT, a Multi-Scale Attention-Guided Hybrid Cervical Transformer framework that combines dense hierarchical convolutional learning and window-based contextual modeling. In contrast to the traditional convolutional networks that mainly learn the local texture patterns, or the transformer-based networks that only focus on the global attention mechanism to improve the model, MAH-CerviT is a synergistic network with three major components: (i) a multi-scale dense convolutional backbone to learn the hierarchical morphological feature aggregation, (ii) a sequential channel-spatial attention refinement module to selectively enhance the diagnostically relevant cellular regions, and (iii) a Swin Transformer encoder to model structured contextual dependency. The hybrid fusion approach allows to discriminate between morphologically similar classes like HSIL and LSIL in a balanced manner and keep the computational efficiency. Massive experiments in 5-fold stratified cross-validation show that MAH-CerviT has a total accuracy of 95.92, macro F 1-score of 95.52 and macro-AUC of 96.40, significantly higher than DenseNet-121, EfficientNet-B4, ViT-Base, and Swin Transformer baselines. It is statistically proved that the increase in comparison to the strongest baseline is significant (p < 0.01). Ablation experiments demonstrate that multi-scale aggregation, attention refinement and contextual encoding have a cumulative beneficial effect on the performance of the baseline convolutional model, which justifies the complementary role of each. The robustness analysis also shows that the performance remains stable even when there are image corruptions and less data. The suggested MAH-CerviT framework provides an efficient trade-off between the accuracy of the discrimination system and the computational efficiency, and it is a reliable and scalable tool to solve the automated cervical cancer screening application.