Glossary

annotation

the process of carefully drawing a digital outline around an object using annotation tools

https://kmlvision.atlassian.net/wiki/spaces/KB/pages/3202580481

anchor point

a bending point to modify the outline of an annotation or region of interest

https://kmlvision.atlassian.net/wiki/spaces/KB/pages/3202580481

background

background is the definition of an area where no sample/specimen is visible

https://kmlvision.atlassian.net/wiki/spaces/KB/pages/3202252814

bounding box

a rectangular outline of an object of interest, defining its coordinates and serving as a point of reference for object detection

confusion matrix

a table chart used to evaluate the performance of a computer vision AI application, where the actual observations are compared to those of the AI model. The matrix offers an overview of correct and incorrect predictions made by the model. The table lists the four values needed for calculating performance statistics: (1) true positives, (2) false positives, (3) true negatives, and (4) false negatives. A well-performing application should have higher values along the main diagonal of the confusion matrix (true-positive to true-negative) and low values on the other positions (false-positive to false-negative).

false-positives

a detection error where the trained application falsely registers the presence of an object

false-negatives

a detection error where the application falsely registers the absence of an object

ground truth

the annotations in the input image data done by the user. The deep learning model is trained to target morphological structures in novel image data, similar to the ground truth annotations.

image classification

a computer vision task for categorizing groups of pixels or vectors within a given image

instance correctness

a metric that gives you information about the ability of the application to correctly detect and label instances in the process of instance segmentation training. An instance is shown as correctly detected, if the predicted instance has an area overlap of ≥ 50% with the annotated instance.

label, labeling, label category

a label is a verbal categorization assigned to an object in a particular image upon annotating the object

multichannel images

images that consist of different channels taken under different laser light conditions that are stacked one above another

multidimensional microscopy images

images visualizing the dynamics of cellular structures in space and through time. These can be three-dimensional (3D), four-dimensional (4D), multichannel or time-series images etc. Multidimensional images are acquired with optical sectioning techniques or laser-scanning microscopy techniques, time-lapse or multifocal-planes recordings. Such images contain additional layers of information on cell movements, structural dynamics, multichannel spectrums and wave-lengths.

object detection

a computer vision task for identifying objects within a digital image

pixel correctness

a metric that gives you information about the correctness of the prediction of the trained model on a pixel level. For each pixel it shows if the predicted pixel label is correct.

predicted areas

areas within an image assigned to a particular label

predicted instance

an individual object that has been detected, segmented and classified by the instance segmentation model

predicted pixel label

the label the AI application assigns a pixel to. This means each pixel in an image is assigned a class label from a predefined set. In other words, the “predicted pixel label“ is the label the AI application assigned to a specific pixel.

random split

the process of letting the platform decide which images are used for training and which ones for validation.

RGB images

images that consist of 3 channels: red, green and blue.

semantic segmentation

a computer vision task for segmenting an image through assigning a label to each pixel within.

time-series images

a set of images taken from the same specimen and with the same modalities but at different times

training images

the images that are used as visual data input to train a computer vision model

true-negative

a prediction whereby the application has successfully recognized an object as not belonging to a class

true-positive

a prediction whereby the application has successfully recognized an object as belonging to a class

unweighted computing

a mode of computation where each visual input or label has equal relative importance within the calculation

validation images

the images within a testing dataset used in order to evaluate the performance of a computer vision model

WSI

Whole-Slide Imaging (WSI) is the process of acquiring high-resolution digital pathology images through the scanning of conventional glass slides