Skip to content

demo_data_tutorial.md

Latest commit

 

History

History
120 lines (73 loc) · 4.66 KB

demo_data_tutorial.md

File metadata and controls

120 lines (73 loc) · 4.66 KB

Demo Data and Basic Usage

This document introduces the demo datasets and the basic usage of SPROUT.
We demonstrate how to use seed generation, adaptive seed generation, growth, and Sprout+SAM.

For more advanced scenarios and parameter tuning, see scenarios.md.

For the napari quickstart, see napari Quickstart Guide.

Specs

Windows 11, equipped with an 11th Gen Intel(R) Core(TM) i7-11800H CPU @ 2.30 GHz, 48 GB RAM.

Demo: Dog Veterinary CT, Target = Bones

Dataset

  • File: dog_img.tif (resampled CT volume)
  • Shape (z, y, x): (691, 161, 122)
  • Intensity range: 0–255 (8-bit)

We can understand that using a threshold range of (120, 255) (using preview function in napari-sprout) yields a relatively complete reconstruction of the skeleton.

Example: simple seed generation

To segment four specific bones mentioned in the manuscript (scapula, humerus, radius, and ulna), we tested several thresholds — (210,255), (220,255), (230,255) — with erosion values 0 or 1, segments = 8, and different erosion footprints (ball, XY disk, XZ disk, YZ disk).

A configuration file is provided in ./template/make_seeds.yaml. Run with:

python sprout.py --seeds --config ./template/make_seeds.yaml

The process finishes in under one minute (3 threads).

From the results, the best setting was:

  • threshold = (220, 255)
  • erosion = 0
  • segments = 8

This produces eight connected components, including the four target bones and extra parts. The final cleanup can be done interactively with napari-sprout->sprout_edit.

Example: threshold-based adaptive seed generation

Instead of using a single threshold, we can define a list threshold ranges, for example lower thresholds from 120 to 220 with a step of 10.

The output will segment the four target bones, but allows a more flexible parameter selection (i.e., instead of a fixed single threshold).

Also if the task requires segmenting more bones, adaptive seed generation can retain small parts.

Example: Threshold-based Growth

Growth can be applied by with multiple threshold ranges (from narrow to wide). This allows seed regions to expand incrementally and helps control overgrowth.

The example seed can be found in: ./demo_data/dog_seed.tif

A configuration file is provided in ./template/make_seeds.yaml. Run with:

python sprout.py --grow --config ./template/make_grow.yaml

In this example, we use 6 threshold ranges, lower threshold decreasing from 220 to 120:

  • First 4 ranges: 20 dilation steps each

  • Fifth range: 5 dilation steps

  • Final range: 1 dilation step

The process finishes in around 20 seconds (4 threads).

As a rule of thumb, we recommend using larger dilation steps in the early stages ( early stopping can prevent unnecessary computation), ensuring components grow to similar sizes at each level. In later stages, fewer dilation steps are sufficient to refine fine details.

Demo Data: Foraminifera Binary Segmentation

Dataset

  • File: foram_img.tif
  • Shape (z, y, x): (215, 1015, 992)
  • Intensity range: Binary (0, 255)

Erosion-based Adaptive Seeding

The goal is to segment individual chambers. This demo image contains 18 chambers in total.

  • Because the input is binary, a direct threshold only produces a single region.

  • A preview without erosion yields 6 connected regions, with many chambers fused together.

  • Increasing erosion helps separation, but excessive erosion causes chamber loss. For example, at erosion = 5, most chambers are separated, but only 13 regions remain because smaller chambers are eroded away.

A configuration file is provided in ./template/make_seeds.yaml. Run with:

python sprout.py --adaptive_seed --config ./template/make_adaptive_seed.yaml

The process finishes in around 8 minutes (4 threads).

The final cleanup can be done interactively with napari-sprout->sprout_edit.

Single-threshold Grow

Since this dataset is binary, a simple grow operation can be applied with single threshold. Using grow iterations = 10, ensuring that all chambers are fully expanded back after erosion.

The process finishes in around 2 minutes (4 threads).

Demo Data: 2D Cell Microscopy Image

Data: 2d_cell.tif

  • Shape (y, x): (2796, 2796)

Example: Use SAM-based inference

In this example, we use 2d_cell_seed.tif as the input seed mask to generate prompts.
The configuration file is provided at ./template/sam_predict.yaml.

Run the following command:

python sprout.py --sam --config ./template/sam_predict.yaml