LABEL (Array-seq): Whole-body molecular and cellular mapping of the laboratory mouse

## LABEL: Spatially Aware Histology-Based Classification **LABEL** is a pipeline that integrates histology feature extraction with a spatially aware hierarchical classifier to predict **organ**, **tissue region**, and **cell type** identities from H&E-stained images. --- ### Pipeline overview #### Preprocessing and feature extraction H&E images are stain-normalized and segmented into image patches centered on spatial transcriptomic capture spots. Each patch is processed using a pretrained **UNI pathology foundation model** to generate fixed-dimensional feature embeddings. #### Hierarchical prediction Predictions are performed using a spatially aware **k-nearest neighbor (kNN)** classifier that incorporates image features and spatial coordinates. Independent classifiers are trained for each hierarchical annotation level: - Organ - Tissue region - Cell type #### Training and evaluation schemes - **In-distribution evaluation** Models are trained on a fixed subset of spatial spots and evaluated on held-out data from the same sections. - **Leave-one-section-out evaluation** Models are trained on multiple whole mouse-sections and evaluated on a fully held-out section. #### Performance assessment Model performance is summarized using **normalized confusion matrices**, restricted to classes consistently observed across replicates.