Title: Integration of DomainSpecific and DomainIndependent Ontologies for Colonoscopy Video Database Annota
1Integration of Domain-Specific and
Domain-Independent Ontologies for Colonoscopy
Video Database Annotation
- Jie Bao1, Yu Cao2, Wallapak Tavanapong2, and
Vasant Honavar1 - Artificial Intelligence Research Laboratory1
- Multimedia Research Laboratory2
- Department of Computer Science
- Iowa State University, Ames, IA USA 50010
- baojie, yucao, tavanapo, honavar_at_cs.iastate.edu
2Outline
- Background of Colonoscopy Video
- Ontology for Video Annotation
- Construction of Domain Ontology and Video
Ontology - Discussion
3Colonoscopy
- Colon is part of intestine and Colonoscopy is an
important endoscopic screening modality for
colorectal cancer - Colonoscopy allows inspection of the entire colon
and provides the ability to perform a number of
therapeutic operations (e.g., polyp removal) - A sequence of images of the colon generated
during endoscopic procedures provides invaluable
information for colorectal research.
4Colonoscopy Analysis System
Semantic units of the colonoscopy video, such as
scene and operation shots, are recognized by
automatically parsing video and audio contents.
Background work The experimental system for
parsing colonoscopy videos into semantic units
5Outline
- Background of Colonoscopy Video
- Ontology for Video Annotation
- Construction of Domain Ontology and Video
Ontology - Discussion
6Semantics in Colonoscopy Videos
- Integration of
- Multimedia Description
- Structure Units Scenes, Shots
- Visual Features Color, Texture
- Audio Features Dictations of the endoscopist
- Domain Description
- Reasons, Procedures, Findings, Diagnosis,
Anatomical Sites, etc.
- Useful for
- Standard description for diagnosis
- Automated retrieval and inference
- Data-driven knowledge acquisition, such as
machine learning
7Semantic Units in Colonoscopy Videos
- View is part of a video for one continuous
inspection - A scene corresponds to an important part of
- the colon such as Rectum and Sigmoid
- An operation shot corresponds to a biopsy or
therapeutic operation such as polyp removal
View
8Colonoscopy Video Ontology
- Those semantics could be represented by Ontology
- Ontology provides a formal conceptualization of a
given domain - In the context of knowledge modeling, an ontology
is a description of the concepts and
relationships of interest - Example Frame isPartOf Shot Congested equals
Edematous
9Ontology Design Considerations
- Expressiveness Sufficiently expressive to permit
precise specification of the entities of the
domain and the relations and constraints among
the entities - Formal with precise syntax and semantics
- Inferentially adequate Permit automated
inferences - Example Enable an agent to infer that if a Video
is about Colon, all scenes in that video are
likely to be about Colon. - Unified framework Describe both the structure of
the video as well as domain-specific terms and
relationships among terms within a single unified
framework
10How to represent OntologySyntax Level
- XML
- Examples MPEG7, Dublin Core
- Has a limited utility as a knowledge
representation language - Semantic Ambiguity
- partOf or subClassOf ?
- Query over syntax, not over semantics
- XQuery
11How to represent OntologySemantic Level
- Ontology Language
- Use DAMLOIL or OWL (Web Ontology Language)
- Extension of RDF(Resource Description Framework)
- Usually use XML as syntax
- Standard technique used in Semantic Web
- Ontology entities class, property, instance
- Class (concept) Colon subClassOf Site
- Property (role) domainVideoClip hasID
rangeinteger - Instance () Video_010 typeVideo
- Can be mapped to description logic, enabling
automatic and tractable inferences - DL is a subset of first order logic
- Inference engines available FACT, RACER
12Outline
- Background of Colonoscopy Video
- Ontology for Video Annotation
- Construction of Domain Ontology and Video
Ontology - Discussion
13Domain Ontology - MST
- Minimal Standard Terminology (MST) a controlled
vocabulary for colonoscopy and endoscopy
reporting - The terms and their relations in MST
documentation are given in a set of tables - Class in taxonomy (subclass) of terms, like site,
findings, reasons, and diagnosis. - Terms listed in the tables as Attributes and
Attributes Values are modeled as properties of
corresponding classes
14Domain Ontology - Part
15Video Ontology
- Extend the set of MPEG-7 descriptors
- Semantic units organized in partOf hierarchy
- hasStart, hasEnd
- Controlled inheritance of properties - common
properties like hasSite, and hasID are modeled in
VideoClip - Includes video and audio features
16Video Ontology
17Integration of Domain Ontology and Video Ontology
- Instances of MST ontology are added as features
to the instances of video ontology - Annotate different levels of video ontology with
different MST annotations - Example Reasons and Examination properties are
only meaningful at the Video level, but not at
the Frame or Shot levels - Cardinalities Multiple annotation of them are
legal
18Integrated Ontology for Video Annotation (IOVA)
19Integration of Domain Ontology and Video Ontology
(Part)
20Example instance of this ontology
21Example instance of this ontology (Part)
22Example instance of this ontology (Part)
23Outline
- Background of Colonoscopy Video
- Ontology for Video Annotation
- Construction of Domain Ontology and Video
Ontology - Discussion
24Conclusions
- In this paper, we present an approach to
annotation of a video database using a domain
specific ontology, a domain independent video
ontology that encodes the structure and
attributes of video data. - The two ontologies are integrated using
domain-specific semantic linkage. The result
(IOVA) is represented in OWL. - Annotation of video data using IOVA constitutes
an important first step towards flexible and
fully automated indexing, retrieval, inference,
and data-driven knowledge discovery using video
data in a broad range of applications including
colonoscopy video analysis.
25Questions
26VIDAI -video database annotation using IOVA
27VIDAI
- Video analyzer Do semantic segmentation and
labeling based on video and audio processing. - Domain Term Annotator Semantic finding from the
video analyzer is created as instance of MST
terms and be annotated with appropriate video
entities. - Ontology Editor and Browser expert knowledge
about the inspected video, such as examination,
complications, reasons, finding. Diagnosis and
therapy, could be added into the video case,
resulting in fully annotated video - Video Repository Annotated video is saved to the
video repository, with support from some backend
DBMS. - Query Interface an interface that could be
visited by agent or expert. Could based on RQL or
RDQL - Browsing interface A interface for user to view
the query result