Use Case 3: Image Feature Variability and Its Impact

The task in this use case is to quantify the image feature variability. We will analyze numerical variability of image features extracted using multiple feature extraction software packages. The use case presented here was selected to shed some light on the importance of measurement science.

Challenges

Can we obtain the same numerical values of image features in multiple labs? There are many ways to approach this question since the number of potential error sources is very large across laboratories. One of the approaches is to investigate the use of different software packages for computing the same image features. We will assume that all other variables in multiple labs would not contribute to the image feature variability.

Inputs

We will be working with the live phase contrast 3T3 images comprised of 238 images and a total of 8162 cells with different shapes and sizes.

Image Processing Pipeline

The pipeline to extract image features is identical to use case 2 and consists of:
  • 1. creating a new collection and uploading images,
  • 2. creating a stitching vector,
  • 3. intensity scaling and pyramid building,
  • 4. segmentation,
  • 5. tracking, and
  • 6. feature extraction.
WIPP image features extraction pipeline
WIPP image features extraction pipeline

Select all software packages when computing shape features to compare their respective values. The output from WIPP will be a csv file of selected features for all tools. The user may explore this output as described in [reference for conference paper and book].

Normalized error between Theoretical value of the perimeter and the measured value by the software package over simulated images of a circle with different diameter size.
Normalized circumference error vs circle radius
Normalized circumference error vs circle radius

The tools agree amongst each other for large radius values, but there is always a bias between the computed perimeter values and the reference value regardless of the software package. This analysis shows that it is also not accurate to analyze datasets that contain small regions of interest, e.g. objects with radius less than 10 pixels or an equivalent area of 350 pixels or less. Higher acquisition resolution is recommended for for feature extraction over small regions of interest.