Multivariate analysis of MRIMRSI data for newly diagnosed glioma patients

1
Multivariate analysis of MRI/MRSI data for newly
diagnosed glioma patients

• Xiaojuan Li
• PhD Candidate, UCSF/UCBerkeley Bioengineering
• 3rd UC Biomedical Symposium
• May 06, 2002

Magnetic Resonance Science Center Department of
Radiology University of California, San Francisco
2
Goal of the project
Combining MRI/MRSI to improve diagnosis for newly
diagnosed glioma patients

• To study relationship of volume and spatial
  distribution between morphologic and metabolic
  abnormalities
• To develop a multivariate analysis method
  integrating information from MRI and MRSI to
  characterize the lesion

3
Materials and Method

• 49 untreated glioma patients, classified by
  histology
• 15 grade 2, 17 grade 3 and 17 grade 4

4
Metabolic index images
13
CNI
ChCrI
0
CrNI
LLI
5
Power of MRI on grading
Grading criteria based on morphologic information
Classification errors with grading method above
Total error 0.204
6
Volume of Metabolic abnormalities
CrNI2
CNI2
(cc)
(cc)
80
50
40
60
30
40
20
20
10
0
0
2
3
4
2
3
4
LLI5
ChCrI2
(cc)
(cc)
30
60
20
40
10
20
0
0
2
3
4
2
3
4
7
Classification tree with morphologic information
Cross validation error rate 10/49 20.4
8
Recursive Partitioning Tree (RPT) and Linear
Discriminant Analysis (LDA)
X

• Apply LDA to subgroups, obtain LDA function
• Apply this function to all of the observation and
  add this resulted vector to the tree construction

9
Classification tree with MRI/MRSI parameters and
LDA function
Y
N
NecL
N
Y
CEL
G4
lt4.2
lt10.1
gt4.2
gt10.1
LDA
Med CNI
G4
G3
G3
G2
Cross validation error rate 5/49 10.2
10
Grade 4 lesion w/o necrosis vs.
enhanced grade 3 lesion
Grade 4
Grade 3
11
Full tree structure
NecL
CE
G4
Med CNI
LDA
G4
G3
G3
G2
12
Conclusions and discussions

• Recursive partitioning tree is a powerful tool to
  develop strategies for combining the information
  from MRI and MRSI for non-invasive grading.
• Easy to interpret
• Easy to combine with other method
• Easy to integrate more variables
• Exploratory tool and the results in this study
  can be data dependent because of the small
  patients population
• Future studies will evaluate the classification
  scheme in a prospective cohort of patients.

13
Fun with tree structure!