BBBC038 Segmentation Benchmark Workflow

This demo workflow compares several nuclei segmentation methods on BBBC038-style nuclei images and evaluates each predicted label image against reference instance masks.

Goal

Benchmark Cellpose v3, Cellpose-SAM, StarDist, and a classical threshold method on a shared nuclei segmentation task.

Data

The workflow follows the BBBC038v1 stage1_train layout. Each sample folder contains one raw microscopy image under images/ and one or more per-object reference masks under masks/.

How It Works

The workflow lists sample folders, combines the per-object mask files into one reference label image, prepares a 2D intensity image for segmentation, runs each method on that image, and writes one benchmark row per method and image.

Results

• Predicted label images from Cellpose v3, Cellpose-SAM, StarDist, and a classical threshold method.

• Overlay previews for visual comparison.

• A benchmark table with object counts, foreground IoU, and Dice scores.

Run

python example_workflows/bbbc038_segmentation_benchmark/workflow.py