Logo: to the web site of Uppsala University

Cytological examination plays an important role in cancer screening and diagnosis by enabling cellular abnormalities to be assessed through minimally invasive sample collection. The increasing digitization of cytology specimens creates opportunities for computer-assisted image analysis, but also exposes fundamental methodological challenges. Abnormal or malignant cells may be extremely rare among vast numbers of normal cells, detailed cell-level annotations are costly and difficult to obtain, and cellular appearance varies across specimens, acquisition conditions, transformations, and imaging modalities.   

This thesis comprises five papers that develop monomodal and multimodal image-analysis methods for abnormal-cell detection in computational cytology under limited supervision. The first part investigates representation learning from normal cells as an alternative to conventional weakly supervised learning. Self-supervised and deep one-class methods are trained using cells from slide-negative specimens and subsequently used to rank cells according to their deviation from learned representations of normality. The results demonstrate that one-class learning can improve the retrieval of rare abnormal cells, particularly at very low witness rates where slide-level supervision provides only a weak and noisy learning signal.

The thesis further extends normality learning in several directions. A symmetry-aware diffusion framework incorporates rotations and reflections that preserve cellular identity, producing more consistent reconstructions and more stable abnormality rankings. Multimodal one-class learning is investigated on the BSCCM and CHAMMI WTC-11 datasets using complementary label-free and fluorescence-derived measurements. The results show pronounced differences in modality informativeness and demonstrate that multimodal fusion can improve upon the strongest individual modality, although the most effective fusion mechanism depends on the imaging configuration and on how cross-modal information is exchanged. The thesis also examines how available slide labels can be used more effectively through slide-label-aware multitask pretraining, where different objectives are applied to negative- and positive-bag patches and adaptive gradient surgery reduces interference between the learning tasks.

Collectively, the findings show that learning appropriate representations of normal cells, incorporating biologically meaningful image symmetries, exploiting complementary imaging modalities, and using weak supervision according to its reliability can improve the retrieval of rare abnormal cells. The results also show that multimodal learning is not uniformly beneficial: its effectiveness depends on modality informativeness, fusion design, and optimization stability. These developments support interpretable, human-in-the-loop computational cytology systems in which a manageable set of suspicious cells is prioritized for further assessment by medical experts.