Whole-Slide Prediction Basics • petrographer

This vignette walks through a complete whole-slide inference workflow using petrographer. It mirrors the experience of a user who has a scanned thin section, wants to run the pretrained detector with SAHI slicing, and then review both the visual overlays and quantitative statistics.

Load the Toolkit

The package ships with all high-level helpers. Load it alongside a few tidy utilities you may want for inspection.

library(petrographer)
library(dplyr)
library(fs)

Point to Your Whole-Slide Image

Set the path to the slide you want to analyse. Replace the placeholder with your .svs, .tif, or other supported image file.

wsi_path <- system.file("extdata", "test_image.jpg", package = "petrographer")
file_exists(wsi_path)

Swap in your .svs, .tif, or other supported whole-slide file once you are ready to run the full workflow.

Cache a Pretrained Model

Pretrained weights are hosted on the petrographer hub. Install (or refresh) the pin you need, then instantiate the R-side wrapper. Replace model_id with whichever detector matches your slide.

model <- from_pretrained('inclusions_shell')

pg_install_pretrained() downloads the pin once and reuses the cache next time. The returned PetrographyModel retains the manifest so you can inspect provenance or training metadata as needed.

Run SAHI Inference with Custom Strides

Whole-slide images are large, so enabling SAHI slicing keeps detections accurate while fitting within GPU or CPU memory. Control the stride through slice_size and overlap.

output_dir <- path("results", path_file(path_ext_remove(wsi_path)))

predictions <- predict_image(
  image_path = wsi_path,
  model = model,
  use_slicing = FALSE,
  slice_size = 1024,  # pixels per tile; shrink for tighter memory budgets
  overlap = 0.2,     # tile overlap to smooth boundaries
  output_dir = output_dir,
  save_visualizations = TRUE
)

predictions

The tibble contains one row per detected grain, complete with geometric measurements and morphology descriptors ready for downstream analysis.

Visualise the Overlay and Exports

SAHI writes annotated overlays alongside the raw detections. Locate the generated assets and preview them in-line.

dir_ls(output_dir)

viz_path <- dir_ls(output_dir, glob = "*_prediction.png", fail = FALSE)
if (length(viz_path)) {
  image_read(viz_path[[1]])
}

You can also hand the exported visual to image viewers or documentation pipelines — each file is saved with the original slide name plus _prediction.png.

Summarise Detection Statistics

With the predictions tibble in hand, compute per-image summaries or overall population metrics to quantify mineralogy at scale.

per_image <- summarize_by_image(predictions)
per_image

population <- get_population_stats(predictions)
population

These summaries surface object counts, size distributions, and shape metrics that support QA checks or comparisons between slides.

Next Steps

Adjust slice_size and overlap to balance runtime and fidelity for extremely large slides.
Pipe predictions into custom tidyverse workflows for further filtering, plotting, or reporting.
Use predict_images() to batch-process entire slide directories once you are satisfied with the single-slide results.