Visualization and analysis

Tissue samples stained with mIHC were scanned with multispectral imaging microscopy (Vectra 3, Akoya Bioscience). Scanned multispectral images were unmixed on inForm software (ver.2.4.0, Akoya Bioscience) to acquire the fluorescence signal from each marker⁸⁶. Imaging analysis was performed on inForm software by identifying tumor (CK + area) and stroma (CK- area proximal to the epithelium), each nucleated cell, and its cell type. Alternatively, QuPath software 2.3.1⁸⁷ was also used to perform similar imaging analysis on unmixed images converted to multi-layered TIFF format by inForm software⁸⁶. The images of the regions of the same type (DCIS or normal) from the same case, were typically stitched together and stored and shared as one single larger multi-layered TIFF image (data availability below). Scanned image areas were aggregated into up to three histological regions per sample: main pre-invasive lesion, alternate pre-invasive lesion, normal epithelium. In each region, the stromal and epithelial densities of each cell type and state were calculated, including when cells were not present (density = 0). Regularized marker densities into distribution deciles were then used to classify samples using non-negative matrix factorization (Supplementary Table 7c). The immune states were assigned and named after the hierarchical clustering of the H-matrix (meta-marker values).