Abstract
Spatial imaging proteomics modalities, such as imaging mass cytometry, enable comprehensive identification of immune microenvironments driving disease outcomes. Identifying outcome-associated immune microenvironments from these data has proven to be complex, as it requires segmenting cells with complex shapes and reconciling spatial signatures across many heterogeneous samples. We present MICRON , a segmentation-free, fully automated multiple-instance learning based tool for automatic identification of outcome-linked immune microenvironments. MICRON learns representations of samples profiled with spatial imaging proteomics modalities, enabling more accurate prognostic and diagnostic prediction over existing approaches. As a case study, we show that MICRON generates a comprehensive importance map that reveals key outcome-associated immune microenvironments in brain cancer, uncovering coordinated cell-cell communication between astrocytes, NK cells, and macrophages linked to survival outcomes. MICRON is provided as open source software for broad use by clinicians and biologists at https://github.com/ChenCookie/micron .