Information Extraction from Radiology Reports: System Design and Implementation

1
Information Extraction from Radiology Reports
System Design and Implementation

• Information Model
• System Architecture UIMA
• Automatic Report Segmentation
• NER
• Negation Discovery
• Coreference Resolution, Relationship Discovery,
  Inference

2
Information Schemas

• DICOM SR
• AIM (Information and Image Markup)

3
(No Transcript)
4
(No Transcript)
5
(No Transcript)
6
Coding Scheme - Code Value - Description
UMLS-2008AA,C0006141, Breast Body Part, Organ,
or Organ Component.
7
(No Transcript)
8
Overall System Design
9
Automatic Report Segmentation
10
SVM classifier achieved accuracy of over 0.9
11
NER

• Image ROI and Image Referents Discovery
• Imaging Observations and Characteristics
• Imaging Procedure
• Body Parts and Organs
• Findings and Abnormalities
• Persons, Dates, Times

12
Image and Image ROIs Referents
13
NER OBA and MetaMap
14
(No Transcript)
15
(No Transcript)
16
Negation Discovery
17
The NegEx Algorithm

• A rule based system for the discovery of
  negation of ?ndings and diseases in discharge
  summaries.
• A list of 35 negation phrases - negations
  preceding a term (e.g. not signs of, no evidence
  of, negative for ), negations following a term
  (e.g. declined, unlikely ), and what they refer
  to as pseudo negations - false negations
  triggers such as double negatives or ambiguous
  negations (e.g. not necessarily, not rule out,
  not certain wether ).
• A 5-token window around the negation cue

18
BioScope

• Clinical free-texts (radiology reports),
  biological full papers, and biological paper
  abstracts from the GENIA corpus.
• Minimal retrocardiac opacity, ltxcope
  id"X382.1.1"gt ltcue type"speculation"
  ref"X382.1.1"gtlikelylt/cuegt atelectasislt/xcopegt.
• Normal chest x-ray ltxcope id"X394.1.1"gt ltcue
  type"negation" ref"X394.1.1"gt withoutlt/cuegt
  radiographic evidence of residual
  bronchopulmonary dysplasialt/xcopegt.

19
Coreference Resolution

• Coreference resolution is the process of
  determining whether two expressions in natural
  language refer to the same entity in the world.
• The largest lymph node is inferiorly positioned
  in the level IV and measures 29 mm in diameter.
  Just superior to this, there is a necrotic lymph
  node measuring 16 mm in size.

20
Relationship Discovery

• The goal of relationship extraction is to detect
  occurrences of a prespeci?ed type of relationship
  between a pair of entities of given types.
• Associations between an imaging observation and
  imaging observation characteristics.
• Between an imaging observation and a body part or
  organ.
• Between imaging observations/characteristics and
  inferred diagnosis.
• Spatial relationships.

21
Inference Module

• Domain speci?c inference module that would have
  the ability to ?ll in gaps in the relationships
  between the named entities present in the report.
  
• The inference module will also be used to
  validate the output of both named entity and
  relationship discovery modules.

22
Q/A