ActiveVisium: Leveraging Active Learning to Enhance Manual Pathologist Annotation in 10x Visium Spatial Transcriptomics Experiments

Abstract

Spatial transcriptomics (ST) technologies offer valuable insights into tissue organisation by capturing gene expression within its spatial context. Among these, 10x Visium stands out for its capacity to integrate gene expression profiles with histological images, facilitating multi-modal tissue analysis. However, comprehensive analysis requires manual pathologist’s annotations at the capture spot level, a labour-intensive and time-consuming process that demands a significant amount of pathologists’ time. Given the scale of studies involving multiple ST samples, manual annotation becomes impractical, and no automated solutions currently exist. To address this, we introduce ActiveVisium, an active learning framework designed to enhance spot-level annotation in 10x Visium datasets. To the best of our knowledge, ActiveVisium is the first framework to leverage tissue morphology and, optionally, gene expression data to automate large-scale spot annotation while selecting the most informative ones for manual labelling. Furthermore, this approach enables transfer learning across similar samples, thereby reducing annotation time for entire studies. Evaluations across breast cancer, colorectal cancer, and healthy kidney samples demonstrate that ActiveVisium has the potential to significantly improve annotation efficiency and consistency. All code and data are publicly available.