Self-supervised learning enables unbiased patient characterization from multiplexed microscopy images
Multiplexed immunofluorescence microscopy offers detailed insights into the spatial architecture of cancer tissue. However, classical single-cell analysis approaches are limited by segmentation accuracy, reliance on predefined features, and the inability to capture spatial interrelationships among cells. We developed a hierarchical self-supervised deep learning framework that learns spatial protein marker patterns from multiplexed microscopy images by encoding the tissue at both local (cellular) and global (tissue architecture) levels. Applied to lung, prostate, and renal cancer tissue microarray cohorts, our method stratified patients into prognostically distinct groups with significantly different survival outcomes. These groupings were consistent with prior expert-driven single-cell analyses, demonstrating the validity of our approach. Furthermore, attention maps extracted from these models highlighted biologically relevant tissue regions associated with specific marker patterns. Overall, our framework effectively profiles complex multiplexed microscopy images and offers a scalable, interpretable tool for improved biomarker discovery, with potential to support more informed cancer treatment decisions.