Nuclei segmentation tutorial with the DSB2018 dataset

What you will do

This tutorial takes about 30 minutes. You will look through the hosted DSB2018 nuclei set, label nuclei on a few of your own images with the same conventions, and export the result in a format a segmentation model can train on.

Step 1: learn the conventions from DSB2018

Open the hosted copy and page through ten or twenty images before you label anything. The set mixes fluorescence and histology stains, magnifications and illumination quality on purpose, so it shows you the hard cases fast.

Look for three things:

• One shape per nucleus. Every nucleus is its own polygon, even when two touch.
• No overlaps. In the source masks no pixel belongs to two nuclei. Where nuclei touch, the boundary is split between them.
• Edge nuclei count. A nucleus cut off by the image border is still labeled.

Step 2: set up your project and label

1. 1

   Create a project

   Create a project and upload a few of your own microscopy images. Five is enough to learn the workflow.

2. 2

   Create one label

   Go to Ontologies in the sidebar, click New label and name it nucleus. Use a single label. Separate classes for nucleus types are a later decision.

3. 3

   Outline each nucleus with the polygon tool

   Open an image, pick the nucleus label and choose the polygon tool. Click around the boundary of one nucleus and close the shape on its first point. Repeat for every nucleus, including the ones on the border.

4. 4

   Split touching nuclei

   Where two nuclei touch, draw each one up to the shared boundary. Do not draw one shape around both.

5. 5

   Review before you export

   Zoom out and check each image for missed nuclei. Missing labels teach a model that a nucleus is background, which does more harm than a slightly loose outline.

Step 3: export to COCO

Go to Exports in the sidebar and click New schema. COCO is the default format, so name the schema and save it. Open it, click Run export, and when the run finishes use the download button on the Exports list to get the JSON file. Each nucleus comes out as one annotation with its polygon in segmentation.

Step 4: turn polygons into instance masks

Most nuclei models (StarDist, Cellpose, U-Net with a watershed step, Mask R-CNN) train from an instance label image: background is 0 and each nucleus has its own integer. This converts a COCO export with pycocotools:

import numpy as np
from pycocotools.coco import COCO

coco = COCO("export.json")
for image_id in coco.getImgIds():
    info = coco.loadImgs(image_id)[0]
    labels = np.zeros((info["height"], info["width"]), dtype=np.uint16)
    anns = coco.loadAnns(coco.getAnnIds(imgIds=image_id))
    for i, ann in enumerate(anns, start=1):
        labels[coco.annToMask(ann) > 0] = i
    np.save(f"{info['file_name']}.labels.npy", labels)

Later shapes overwrite earlier ones where they overlap. If your labels follow the no-overlap rule from Step 1, that never happens.

Where to go next

• What is inside DSB2018, its folder format and which model to start with: the nuclei segmentation dataset guide.
• A second, more uniform nuclei set from one chemical screen, with an official train, validation and test split: BBBC039 on DataTorch.

Source and license

BBBC038v1, the 2018 Data Science Bowl image set, from the Broad Bioimage Benchmark Collection. License: CC0. Source: https://bbbc.broadinstitute.org/BBBC038. Please cite: Caicedo et al., Nucleus segmentation across imaging experiments: the 2018 Data Science Bowl, Nature Methods 16, 1247 to 1253 (2019).