Newly Diagnosed Glioma patients study with MRI/MRSI by GjnL49L


									                                                               Poster #33

 Integrating in vivo MRI/MRSI data for
 non-invasive classification of gliomas

             Xiaojuan Li, Ying Lu, Andrea Pirzkall, Sarah Nelson

                                ISMRM workshop on
            In Vivo Functional and Molecular Assessment of Cancer
                    Santa Cruz, CA, USA, Oct 19-21, 2002

Graduate Bioengineering Group            Magnetic Resonance Science Center
    University of California                   Department of Radiology
   Berkeley / San Francisco              University of California, San Francisco
                                             Poster #33

Combining MRI and MR Spectroscopic
imaging (MRSI) to improve non-invasive
diagnosis for newly diagnosed glioma
   To explicitly integrate information from MRI and
    MRSI in classification procedure
   To consider MRSI characteristics of the entire
    lesion instead of a pre-selected region from
    within the anatomic lesion.
                                                            Poster #33

   64 untreated glioma patients, classified by histology
       21 grade 2, 22 grade 3 and 21 grade 4
   1.5T GE Signa Echospeed clinical scanner
       T2-weighted images, pre/post-contrast T1-weighted images
       3D MR Spectroscopic imaging
            PRESS, Te=144ms, TR=1s, 1.0 cc voxel size
   Data classified with recursive partitioning analysis
       Morphologic abnormal volumes, metabolic abnormal volumes,
        median, max levels of metabolic abnormal within ROIs, metabolic
        burdens measured by sum of abnormal index within ROIs
                                                            Poster #33
Classification tree with morphologic information
                                                     N NecL     Y
        CEL         NecL
       Y    N      Y        N
 G2    1   20      0       21             N            Y
 G3   14    8      0       22
 G4   21    0     17        4
                                     G2                G3

                CEL: contrast enhancement lesion
                NecL: Macroscopic necrosis

      Cross validation error rate 15/64 = 23.4%
                                                                   Poster #33
Classification tree with MRI/MRSI parameters
V1: burden of Cr/NAA abnormal                     N        NecL    Y
V2:volume of lactate/lipid, volume
ratio of Cho/NAA abnormal/T2, burden
Cho/NAA abnormal

                           N           CEL    Y                         G4

                  V1                                  V2

        G2                 G3                G3               G4

             Cross validation error rate 7/64 = 10.9%
 MRSI helps to distinguish
  Non-enhanced grade 3 lesions from grade 2 lesions
  Grade 4 lesions without macroscopic necrosis from
 enhanced grade 3 lesions
                                     Poster #33
Non-enhanced grade 3 lesion vs. grade 2 lesion

                                      Grade 2

       Grade 3


0       6
                                       Poster #33
Grade 4 lesion without macroscopic necrosis
vs. enhanced grade 3 lesion


                                     Grade 4

           Grade 3


 0          17
                                                        Poster #33
                                   N        NecL

              N      CE        Y

        V1                             V2
  G2          G3          G3                   G4

Full classification tree structure with images of typical
lesions for each branch. The method also provides
clues for possible subgroups within each grade.
                                                  Poster #33
 MRSI provides useful information in addition to
  MRI during non-invasive classification for gliomas.
  It also provides clues for subgroups within each
 Recursive partitioning analysis is a powerful tool to
  develop strategies for combining the information
  from MRI and MRSI for classification.
    With natural and useful interpretation
    Easy to integrate more variables
    An exploratory tool and the results in this study can
     be data dependent because of the small patient
     population. Larger cohort of patients will be studied
     in the future.

This research was supported by NIH R01-CA79719
                                                             Poster #33
Recursive Partitioning Analysis (RPA) &
Linear Discriminant Analysis (LDA)


                  (a)                                (b)
   Both of the algorithms maximize between group differences while
    minimize within group differences. But RPA is a sequential binary
    process and LDA is a one-step linear separation. (a) Classification
    tree; (b) Linear discriminant analysis with discriminant function.
 So our strategy:
  Apply LDA to subgroups, obtain LDA function
  Apply this function to all of the observation and apply this resulted
    vector in addition to other variables to construct the tree
Metabolic index images


                     CNI                ChCrI


                     CrNI                 LLI
  CNI2   CrNI2 Define metabolic abnormalities as regions within
  LLI5   ChCrI2 CNI2, CrNI2, ChCrI2 and LL5 respectively

To top