Annotation Ontology

Arising from initial meeting on 15/3/06 with:

• SusanAstley^?
• TimCootes^?
• ChrisRose^?
• ChrisTaylor^?
• DanieleTuri
• AlanWilliams
• TomosWilliams^?
• KatyWolstencroft

Minutes

• We need to consider 'has a' relationships, e.g. 'face has a nose'.
• Regions of interest (ROI, aka a boundary, called GeometricObject^? in the diagram): can contain interesting features (and non-interesting things) or can define exactly where the interesting feature is. Usually, the former case arises when there is uncertainty about the location/presence of the interesting feature, or where it is time-consuming annotating each feature.
• There is a difference between the surface of a boundary and the volume defined by such a surface.
• We defined 'uncertainty' to mean ambiguity in the presence or classification of a feature.
• We defined 'accuracy' to mean the ambiguity in the exact location of a feature.
• We talked about the situation where a feature is only known about due to its status in another image (e.g. an asymmetry in a pair of mammograms). The feature may only be annotated for a subset of the relevant images. This subset of annotated images may be chosen based on the status of the feature in those images. As a concrete example, an asymmetry in a mammogram would be annotated in the mammogram where the asymmetry is brighter (where the tissue is more dense) because more dense tissue can be a cancer; in dementia, an annotation may be made to indicate the lack of a structure or its dimished size.
• We asked the question "are annotations always measured in the image frame or the world frame", i.e. is there always a transformation made to the image and boundary geometries to normalise them? The answer we arrived at is that this is not typically done and that things are measured wrt the image frame. It was suggested that ChrisTaylor^?, TimCootes^? and ChrisRose^? meet to discuss this further.

The following points refer to the diagram below.

• Accuracy is described in the diagram as being a vector of reals; in reality, the situation is more complex and one would use a model of the probability distribution associated with the parametrisation of the GeometricObject^?. For example, this might be the name of the distribution and the parametrisation of it.
• CertaintyDegree^? describes how certain some process (e.g. expert annotation) is about the presence or classification of the structure. This might be a person's belief expressed on some interval (e.g. [0,1]) or a Likert Scale; or it could be information about a set of competing hypotheses (e.g. Likert scores for each competing hypothesis); or it could be some other type of numerical and/or categorical value(s).
• Creator may be a person or a process (and an output from a process potentially may result from the input of other creators).
• Dimensions is not simply a size; there are many ways in which a geometrical object may be represented (e.g. a pixel map, single points, a set of nodes and arcs, parameterised lines/curves/surfaces/volumes etc.).

Diagram

• Source code for diagram (.dia format)