On this path

Install Python
Set up a working interpreter on your local computer.
Virtual environments
Isolate project dependencies with venv or conda.
Python basics
Learn variables, data types, loops, and functions with exercises.
Scientific Python libraries
Meet the core tools used in research computing and image analysis.
Jupyter notebooks
Run code interactively with examples and exercises.
pandas and matplotlib
Load tables, summarize data, and make plots that tell a story.
napari + mAIcrobe
Launch a viewer workflow for microbial image analysis.
Create a package
Turn scripts into a reusable Python project.
ML segmentation
Use scikit-learn for unsupervised and supervised segmentation.
Cell classification
Train a classifier on single-cell measurements.
Segmentation QC
Filter out poor segmentations automatically.
Neural networks
Build intuition for network structure with NumPy.
Pretrained models
Explore Cellpose and model-zoo workflows.

Module 11

So you want to filter out bad segmentations

This lesson adapts `3_sklearn_segmentation_filtering.ipynb`, which combines instance segmentation, morphological measurements, and classification to identify badly segmented cells.

Estimated time: 75 to 105 min Main tools: scikit-image, SciPy, scikit-learn

Main technical steps from the notebook

Load phase-contrast and binary-mask images.
Fill holes in binary masks.
Create a distance transform.
Find local maxima as markers.
Use watershed to split touching cells into instances.
Measure morphological features.
Train a classifier to detect badly segmented objects.

Why this notebook is especially useful

It teaches a realistic workflow problem rather than an abstract one. In real projects, segmentation errors are common, and a large part of analysis work is deciding which objects to trust.

Website teaching structure

Part 1: build the instance segmentation

Start from the binary mask and show why touching cells are a problem. Walk through hole filling, distance transforms, local maxima, and watershed one step at a time.

Part 2: measure morphology

Once individual objects exist, connect them to per-cell measurements. This is a nice bridge back to the earlier classification module.

Part 3: classify segmentation quality

Present this not as “AI fixes everything,” but as a pragmatic way to reduce manual cleanup by identifying likely bad segmentations.

Representative code examples

Load images and fill holes

from skimage.io import imread
from scipy.ndimage import binary_fill_holes

phase = imread("../data/images/wt_phase_contrast.tif")
mask = imread("../data/images/wt_mask.tif")
filled_mask = binary_fill_holes(mask)

Distance transform and markers

import numpy as np
from scipy.ndimage import distance_transform_edt, label
from skimage.feature import peak_local_max

distances = distance_transform_edt(filled_mask)
coordinates = peak_local_max(distances, labels=filled_mask)
coordinates_tuple = tuple(coordinates.T)
local_maxima = np.zeros_like(filled_mask, dtype=bool)
local_maxima[coordinates_tuple] = True
markers = label(local_maxima)[0]

Watershed

from skimage.segmentation import watershed

labels = watershed(-distances, markers, mask=filled_mask)

Important reasoning skill

Learners should be encouraged to inspect failures visually. A classifier that filters segmentations is only useful if the labels used for training reflect the biological and analytical goals of the study.

Exercises worth integrating

Display the phase image and mask side by side.
Show the hole-filled mask and explain what changed.
Display the distance transform and marker image.
Run watershed and inspect where it succeeds or fails.
Use measured features to train a classifier for segmentation quality.

Back to cell classification Continue to neural networks