ARTICLE

Definition of Image Annotation, Labeling, Segmentation, and Ground Truth Terms

It can be confusing in the world of image segmentation, labeling, annotation, and ground truthing when so many similar terms are used to describe activities and data. In this article we’ll help you understand what these terms mean so you can gain a clear understanding of their use and what each term represents.

Understanding the language of image annotation, object labeling, and image segmentation can enable you to work more effectively with your team. It’s easy to use the terms incorrectly and our goal here is to help you learn what each term means and how to use it properly.

Segmentation Definition

A segmentation is a group of pixels (or voxels in 3D) that represent an object of interest. A segmentation is distinctly different from a bounding box region which merely identifies the rough location of an object within an image. A segmentation identifies specifically which pixels (or voxels) constitute an object. Further, the shape of the segmentation is entirely defined by the underlying object for which it represents. This contrasts with, for example, a bounding box region which contains pixels (and voxels) from both the object of interest and other items surrounding the object. A segmentation is the most precise image-based representation of an object delineation.

Object or Target Definition

An object is the thing of interest or the target you are looking for in an image. For example, a person in a photograph or a tumor in a CT scan. The object is represented by a group of pixels (or voxels in 3D) so that it is composed entirely of pixels that are intrinsically part of the object of interest.

Definition of Annotation, Attribute, and Label (as a noun)

An annotation (or label) is information about the object of interest. For example, a person could have the following attributes: height, weight, location, and age. A tumor could have attributes like grade, type, location, or scanner modality used to acquire the image. Labels ultimately represent meta-information associated with the object of interest that explains what the object is to others.

Definition of Annotating and Labeling (as a verb)

The act of annotating (or labeling) is when a person uses software to assign attributes to segmented objects. For example, the software operator assigns labels to segmented tumor objects indicating sarcoma, breast cancer, etc. Or an objects within a photograph can be assigned attributes like: cat, dog, person, running, walking, and so on. This can be a tedious task without great tools.

Ground Truth Definition

This is the object segmentation that represents the gold standard for a particular image (or scan). The ground-truth segmentation is assumed to be the correct pixel (or voxel) grouping for an object of interest. The ground truth segmentation is typically used as the baseline to compare results from machine learning algorithms to determine the quality of that algorithm.

"I'm amazed that with just a few clicks
my objects are found!"

– Letha R., Annotator of Several Thousands of CT Scans

Does segmentation quality matter?

Hard to use tools slow down your team, waste time and money, and cause them to make mistakes. These errors often go unnoticed by your team and find their way into your machine learning training pipeline where your fancy algorithm learns those same mistakes. And it learns the mistakes well because that's what machine learning is designed to do--learn what you give it.

Take a look at some real-world errors below and decide whether these kinds of issues would impact the final algorithm.

The operator slipped up with the mouse and created a segmentation piece that extends well outside of the aorta being segmented. Machine learning algorithms trained on this scan segmentation will learn this mistake.

Left- and right-side are swapped in this one. Imagine a machine learning algorithm that gets left/right wrong and a surgeon that removes the wrong kidney.

YOUR NEW SEGMENTATION PROCESS

Perform 3D Image Segmentation Better

Better segmentation and annotation tools enable your team to work more effectively. See why Stratovan tools are being used by customers world-wide to streamline their ground truthing efforts.

Import

1

Use what you already have. Import your 3D object segmentations from other tools to improve their quality. You'll be surprised what you find when you switch to Stratovan's toolset. The automated importing process makes it easy and helps you find errors.

Create

2

Use Stratovan's world-class 3D segmentation tools to easily create object segmentations. The automatic segmentation tool quickly extracts most anything what you need. The theshold tool pulls out objects by density. And manual tools can be used to clean anything up.

Annotate

3

Easily apply label attributes to thousands of objects with the push of a button. Share label attributes among your team. Standardize your meta-data and organize your team.

Evaluate

4

Find and fix errors in your 3D object segmentations and annotations. How many errors do you have in your training database? Look at thousands of your objects at once and use statistics to find outliers that indicate problems. Quickly load, review, and fix problematic scans with the push of a button.

Export

5

Export your now high-quality object segmentations to be used as the ground truth for your machine learning pipeline. Several output formats are support to meet your every need.