Abstract Background Cytology plays a critical role in cancer screening and diagnostics. However, developing robust deep learning models for cytology remains challenging because of the high morphological variability of individual cells, differences in staining and sample preparation, and the limited availability of large and annotated datasets. While recent histology foundation models have shown remarkable generalization across tissue types, their direct application to cytology tasks remains unexplored, as cytological slides emphasize cellular morphology rather than tissue architecture. Methods We investigate the effectiveness of applying histology foundation models for cytological classification using a weakly supervised Multiple Instance Learning (MIL) approach. We evaluate several pre-trained histological models as feature extractors, alongside two MIL algorithms, and explore different classification strategies. Results Our results demonstrate that leveraging histology foundation models yields superior performance compared to ImageNet-based extraction or training from scratch, particularly on small datasets. We further develop a cytology-specific feature extractor based on DINOv3, which outperforms larger models with higher parameter counts. Conclusions Our findings support the feasibility of cytology-adapted foundation models for whole-slide analysis and highlight their potential within weakly supervised learning frameworks under limited annotation settings.
Martins et al. (Mon,) studied this question.