How Can a Browser Transform Histopathology Image Analysis?

Imagine looking at a single cell under a microscope, but with a computer that can analyze millions of them in seconds. That’s what SASHIMI does: an online tool that uses artificial intelligence to segment cell nuclei in digital histology slides. After the AI outlines each nucleus, SASHIMI turns those points into a spatial map ready for analysis with 27 traditional statistics, area indices, and topological measures. Highlights include calculating cell density, average distance between different cells, autocorrelation indices, and geometric shapes of cell clusters. The developers tested the tool on two datasets: potentially malignant oral disorders (OPMD) and non‑small‑cell lung cancer (NSCLC). In both cases, certain spatial features strongly correlated with patient survival, suggesting that cell arrangement can reveal tumor aggressiveness. SASHIMI is free, runs directly in a browser, and lets researchers worldwide compare results quickly and reproducibly, opening new ways to understand how cells organize in diseased tissues.

SASHIMI is built on a solid theory of point patterns, i.e., sets of points representing the positions of cell nuclei. Among traditional statistics, the most famous is Ripley's K‑function, which measures the density of points within a distance r, comparing it to a random distribution model (Poisson process). Its derivative, the pair‑correlation function g(r), highlights aggregation or repulsion scales more precisely. When nuclei are marked (e.g., with immunophenotypic labels), marked statistics such as the mark‑correlation function assess whether cells of the same type tend to cluster. SASHIMI also integrates area indices like Voronoi tessellation, which partitions space into cells around each nucleus, providing local density and cell shape measures. For topology, the software uses Persistent Homology, a Topological Data Analysis (TDA) technique that captures “holes” and “voids” in the cellular distribution at multiple scales, representing them in persistence diagrams. These diagrams quantify the structural complexity of tumor tissue. Implementation‑wise, SASHIMI is built with WebAssembly to run intensive calculations directly in the browser, avoiding dedicated servers. Input data (x, y coordinates and markers) are loaded in CSV or JSON format, and results are visualized with D3.js, enabling real‑time interactions: zoom, region selection, and heatmap overlays. Reproducibility is ensured by a deterministic computing environment and export scripts for standard formats (CSV, GeoJSON). Moreover, SASHIMI supports integration with deep‑learning pipelines based on TensorFlow.js, allowing model updates without leaving the browser. This all‑in‑one approach makes the tool accessible to labs with limited resources, promoting large‑scale advanced spatial analyses and facilitating the discovery of spatial biomarkers linked to cancer prognosis.