Learning from Normality in Computational Cytology – Swarnadip Chatterjee

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.

In this seminar, I will present an overview of my PhD work on learning from normality for rare abnormal cell detection under limited supervision. The central idea is to learn normal cell representations using reliably normal slide-negative patches, and then identify abnormal cells as deviations from this learned representation of normality. I will discuss the key contributions: a contrastive self-supervised one-class representation-learning approach for rare abnormal-cell detection; an adaptation of Deep Support Vector Data Description for rare cell retrieval, evaluated across controlled witness rates down to 0.05% on a bone marrow dataset and on real oral-cancer whole slide cytology with blinded expert review of the top-ranked cells; symmetry-aware diffusion for transformation-consistent anomaly detection; multimodal one-class learning using complementary imaging modalities; and slide-label-aware multitask pretraining that uses weak supervision according to its reliability. Collectively, the works show that learning robust normality representations, leveraging biologically meaningful symmetries, utilizing complementary imaging modalities, and interpreting weak supervision carefully, can improve rare abnormal cell retrieval, support focused human-in-the-loop expert assessment, and provide a basis for a bottom-up interpretable approach for whole slide cytology image classification by aggregating evidence from individually detected abnormal cells.

About Swarnadip Chatterjee