PixlMap: a generalisable pixel classifier for cellular phenotyping in multiplex immunofluorescence images

We introduce an alternative approach to cellular phenotyping, inspired by the observation that human experts can reliably assign cell phenotypes without precisely delineating cell boundaries. Our method is a user-friendly, deep learning–based phenotyping technique that mimics this human capability by accurately classifying cells using only nuclear segmentation.To develop this approach, we used human-annotated cellular regions as ground truth and implemented a classifier based on the U-Net architecture within a commercially available deep learning image analysis platform, Visiopharm—though the concept is easily adaptable to any deep learning framework. Importantly, the training process requires just a single example of each type of compartmental stain (nuclear, cytoplasmic, and membranous). The resulting model can assign phenotypic classes to nuclear labels, without the need to reconstruct full cell boundaries.This method is both highly novel and broadly generalisable. It delivers accuracy on par with pathologist-level annotations, and out-performs traditional intensity-based phenotyping techniques, offering a robust solution that bypasses the limitations of inaccurate cellular segmentation while maintaining reliable phenotype classification.