Researchers have built a pipeline that looks at a stained tissue slide, labels every cell in it, and then writes up what it sees in plain clinical language.
The system, called SegTME-UNI2, chains three steps together. First, a segmentation model named UNI2-UperHoVer scans routine H&E-stained slides and classifies every cell nucleus into one of six types, using a two-part decoder setup that also separates touching cells from each other. Second, a feature pipeline crunches those labeled cells into 22 measurements covering density, shape, spatial spread, and distance between cell types. Third, a language model, fine-tuned through NVIDIA's BioNeMo platform, turns those numbers into a written narrative describing the tumor's surrounding tissue. The segmentation model was trained in three stages on increasingly large datasets, ending with 1.6 million patches from the TCGA-UT collection.
The interesting part isn't the segmentation model itself, since cell-classification tools for pathology already exist. It's the decision to route quantitative output through a language model for reporting, while keeping every generated sentence traceable back to a specific measured number. That traceability is the whole pitch: instead of asking a clinician to trust an AI-generated paragraph, the system lets them check any claim in it against the underlying cell counts and spatial stats it came from.
The team validated the pipeline qualitatively on lung cancer slides from the IGNITE dataset, sourced from multiple institutions with different staining processes, and reports the phenotype calls held up despite that variability and imperfect segmentation. They've released the pseudo-labeled dataset and model checkpoints publicly, which matters more than the paper's specific numbers: reproducible, open tooling is what actually moves computational pathology forward, not another one-off benchmark result nobody else can rerun.