Smart deep learning application for single cell classification

Identifying specific cell types within diverse populations is challenging, especially when labels are limited or unknown. We present a self-supervised deep learning approach to classify single-cell scattering patterns captured via microfluidics, enabling the detection of rare cells such as circulating tumor cells. Unlike traditional methods like flow cytometry, our technique is faster, simpler, and preserves native cells for downstream analysis. Conventional classifiers are limited by their reliance on labeled data and closed-set assumptions, leading to misclassification of unseen cell types. To overcome this, we developed an open-set classifier that dynamically updates known classes and uses an auxiliary domain to detect out-of-distribution cells. Through instance weighting and feature extraction, the model learns to identify novel classes without prior labels. This method is adaptable to other imaging modalities, such as bright-field and holography. Our work encourages researchers to explore their model's feature space to better understand classification boundaries and uncertainties, ultimately improving performance in biological sample analysis.