index

Seminar slides and workshop material will be made available on our GitHub page:

https://github.com/ImageAnalysis-RockefellerUniversity

Outline

Segmentation
Cell/particle Counting and Tracking
Image denoising

Open-source Softwares
ImageJ/Fiji, QuPath, Napari, CellProfiler, Icy

Commercial Softwares
Imaris, Arivis, Aivia, Huygens, MetaMorph

Machine/Deep Learning frameworks
Weka, ilastik, Labkit StarDist, Cellpose, DeepImageJ, ZeroCostDL4Mic

What is a Raster Image?

Raster images are composed of a grid of pixels
Each pixel contains intensity information
Number of bits (N) determine the range of intensity levels

Image Type	Range of intensity levels (0 to 2^N-1)
8-bit	0-255
16-bit	0-4095
32-bit	0-65,535
RGB color (3 x 8 bits)	0-255 per channel

Segmentation

Identifying object(s) of interest in an image
- cells, nuclei, membrane, transcription sites etc.

Segmentation is usually followed by quantitative analysis of object(s)
- number of cells/nuclei, mean fluorescence intensity, shape etc.

Divide image into areas representing object(s) of interest and background

Segmentation is not an easy task to solve in most practical cases
- Signal variability throughout the image
- Noise, blur and other distortions caused by the imperfect imaging conditions

Segmentation tools

Global thresholding, local thresholding
Image processing filters – Gaussian, Median, Sobel etc.
Machine learning methods (supervised learning) - Weka, Labkit, ilastik
Deep Learning methods - StarDist, Cellpose
3D segmentation - Imaris, Arivis

Segmentation using global thresholding

Human HT29 colon cancer cells, Image from Broad Bioimage Benchmark Collection, Ljosa et al. 2012 Nat Methods

Segmentation using global thresholding

Human HT29 colon cancer cells, Image from Broad Bioimage Benchmark Collection, Ljosa et al. 2012 Nat Methods

Segmentation with local thresholding

Auto Local
Threshold
in Fiji

Human HT29 colon cancer cells, Image from Broad
Bioimage Benchmark Collection, Ljosa et al. 2012 Nat Methods

Segmentation using Machine Learning

Supervised learning - pixel classification using Random Forest classifier

Fiji (Weka and Labkit plugins), ilastik, QuPath, Napari, CellProfiler

Requires orders of magnitude less training data/resources than Deep Learning methods

Hoechst-stained Nuclei, image courtesy of Cherie Au, Giannakakou Lab, Weill Cornell Medicine

Segmentation using Deep Learning

Most accurate methods available for cells/nuclei segmentation
Step 1: Training
Generating a Deep Learning model is resource hungry:
- High-end workstation
- Large amounts of training data (images and annotations)
- Training could take hours to days
- Good programming knowledge required - Python

Step 2: Prediction
Using the model from step 1 to predict the segmentation results :
- A regular laptop is just fine
- Prediction takes seconds to mins
- Little to no programming knowledge required