Deep learning‑based cell type prediction in lung tissue from brightfield histology using CODEX-derived labels

Identifying cell types present in lung biopsies can provide critical information on pathological processes, tissue organization, and organ health, and are valuable in both clinical and research settings. Knowledge of cell types, their distributions, and spatial relationships can assist pathologists and researchers in diagnosis, prognosis, and mechanistic investigations. Multiplex imaging technologies such as Phenocycler^™, based on co-detection by indexing (CODEX) can provide whole-slide spatial maps of protein expression and can accurately identify cell types; however, CODEX shares limitations including high costs of antibodies and reagents, as well as labor-intensive conjugation and validation steps. In contrast, hematoxylin and eosin (H&E) staining is inexpensive and routinely available, offering the potential for scalable cell-type mapping if robust prediction can be achieved directly from histology. In this work, we develop a deep learning pipeline to automatically detect cell types in H&E-stained lung tissue sections by leveraging ground-truth annotations from paired CODEX images. The dataset comprises over 2.3 million labeled cells, segmented from lung tissue sections obtained from multiple donors, and grouped into five broad classes: epithelial, immune, endothelial, stromal, and contractile. We train a DeepLabV3+ semantic segmentation model with ResNet backbone on patches from these annotated slides and reserve a subset for testing. Our model achieved 51.3% balanced accuracy across five classes, which is ~2.5x the random-guessing baseline of 20%. This framework demonstrates the feasibility of approximating multiplexed imaging-based cell type maps from routine histology.