Medical Image Databases

1
Medical Image Databases

• Hemant D. Tagare
• Dept. of Diagnostic Radiology
• Dept. of Electrical Engineering
• Yale University

2
Examples (Card. MR)
(From Carl Jaffe, Yale Univ.)
3
Examples (Card. Echo)
(From Carl Jaffe, Yale Univ)
4
Examples (Spine X-rays)
(From Rodney Long, NLM)
5
Examples (Carpal bones)
6
Examples (Dorsal Fin)
(Collaboration with UTMB and T AM)
7
Why are they different?

• Image content and semantics
• Hard to describe by text best described
  graphically.
• Hard to analyze automatically.
• Rich in geometry.
• Image features
• Complex.
• Evolve with the database.
• Technically difficult to index.
• Queries
• Wider range (Browsing Research).
• Naïve User.
• Low tolerance for errors.

8
Image Semantics
D. H. Hamer, J. Lindsay Jr., Redefining True
Ventricular Aneurysm, Am. J. Card., vol. 64,
1192-94, 1989.
9
Defining Semantics by Images
Aneurysm
Image
Abstract
Features
Size Shape Motion
Grade 0 Normal
Size Shape Motion
Grade 3-5 Aneurysm with deformities
10
Basic Mechanism
Retrieval by similarity
Database
Shape
Feature
Database Image
Abstract
Similarity
Shape
Query Image
Abstract
Feature
11
Image Databases

• Database creation
• Tools for defining images and semantics.

• Retrieval
• Retrieval by similarity.

12
Key Idea
13
Key Idea (contd.)

• In many bio-medical images, most geometric
  information can be calculated in an axiomatic
  fashion from relatively little information.

Thickness
14
Schema
MRI Axial
LA
RA
LV
RV
O
Explicit
Area
Wall
Septum
Features
Axis
Thickness
15
Segmentation
Deformable Templates Active Contours
1.    Tagare H. D., Deformable 2-D Template
Matching Using Orthogonal Curves,'' I.E.E.E.
Trans. on Medical Imaging, vol. 16, No. 1,
pp. 108-117, Feb 1997.
16
Schema Editor
SCHEMeD
S. Ghebreab, M. Worring, H.D. Tagare, and C.C.
Jaffe. SCHEMed a visual database tool for
definition and entry of medical image data. In R.
Jain and S. Santini, editors, Proceedings of
Visual Information Systems , pages 189-196.
Knowledge Systems Institute, 1997.
17
Database Organization
18
Retrieval By Similarity
Range Query
Query feature y Image feature x Find all
images with features that satisfy
S(x,y) lt T
Nearest Neighbor
Find k images with most similar features.
19
Features, Similarities, Indexing Trees
Features
Vector
Non-Vector
Data partition Cluster trees, Vp-trees, etc.
Metric
Feature Space partition (kd-trees, R-trees etc.
Similarity
???
Non-Metric
Indexing Sub-linear complexity Typ. O(log n)
Tagare H. D., Efficient Retrieval in Image
Databases, S.P.I.E. Conf., San Diego, July 2000.
20
Vectors-Metrics
Vector Area, Axis, Rel. Pos.,
Similarity
Kd-tree
Feature Space
21
Non-vector- Metric
Non-vector Shapes of point-landmarks.
Similarity
Procrustes Distance
Cluster Trees
22
Vectors-Non-metric
Vector Image Histograms, Textures Properties
of Curves
Similarity Intersection distance Chi-square
distance Dynamic Prog. distance
Tagare, H. D., Efficient Retrieval Without
a Metric,'' I.E.E.E. Conf. Visual'97, San Diego,
1997.
23
Cardiac Ultrasound Database
24
Contd.
25
Dolphin Database
26
Contd.
27
Interesting Open Problems

• Curse of Dimension
• Indexing tree performance degrades with
  increasing
• dimension (gt 10).
• Browsing
• Other modes of retrieval besides range and
  Nearest- neighbor.
• User Feedback

28
The Curse of Dimension
29
Optimal Covering
Node Elimination
Provably optimal w.r.t. avg. cost of
retrieval for a range query. For small d,
savings 30. For large d, only leaf
nodes.
  Tagare H. D., Increasing Retrieval
Efficiency by Index Tree Adaptation,'' I.E.E.E.
Workshop on Content-based Access of
Image and Video Libraries, San Juan, 1997.
Qian X., Tagare H.D., Optimally Adapted Indexing
Trees for Medical Image Databases, Intl. Symp.
Biomed. Imag, Washington D.C. 2003
(Submitted).
30
Non-Uniform Data
Most image data is not uniform Can we exploit
this to improve indexing?
Index using intrinsic dimension?
31
Browsing
Move along a path
How can we construct indexing structures for this?
32
User Feedback
Similarity
Procrustes Distance
Is there a way of providing feedback so that the
system is used effectively?
33
Acknowledgements

• C. Carl Jaffe, Yale University.
• James S. Duncan, Yale University.
• Gil Hillman, UTMB.
• Bernd Wursig, TAM.
• Rodney Long, George Thoma, NLM.
• Glynn Robinson, Eric Bardinet.
• Xiaoning Qian, Zhong Tao.
• Supported by the National Library of Medicine.