MAPPING ABNORMAL SUBCORTICAL BRAIN MORPHOMETRY

IN AN ELDERLY HIV+ COHORT

Benjamin Wade , Victor Valcour , Lauren3 Wendelken-Riegelhaupt
,
Pardis
Esmaeili-Firidouni
,
4
1
Shantanu H. Joshi , Yalin Wang , Paul M. Thompson
1

Imaging Genetics Center,

USC
Memory
and
Aging
Center,
UCSF
Ahmanson-Lovelace
Brain
Mapping
Center,
UCLA
4
School of Computing, Informatics, and Decision Systems Engineering, ASU
2

Questions

Results
Shape-based differences
Radial Distance

Volumetric differences
Jacobian Determinant

Accumbens

Volume (scaled) mm3

-4

-4

Basal ganglia, ventricular and white matter abnormalities are commonly

reported in HIV+ cohorts.

2
4 p = 0.15
= 37


p = 0.65
= 27



































Accumbens




 


 

 




 

  


 


Volume (scaled) mm3

Nadir CD4 Count

0
-3

0
-3





















































p = 0.15
= 76





p = 0.019
= 440







































































p = 0.075
= 270































































Control

Illness Duration



 

 


Volume (scaled) mm3

Illness Duration

  

 

 






 
 


    
    
 


  

   











 






    

 
  

   






 








    

 



  

  
 
 















 





  

     


 

 









 


3rd Ventricle







   



     









 
   









p = 0.97
= 0.22







p = 0.97
= 0.58





  
 
  





 




 

 
 

 








p = 0.97

= 0.37





 


   

 

  
 



    

 

 
 

 


 






  


   

 
  



  



 





p = 0.97
= 5.5




 






 
   
 
     
    





 








p = 0.97
= 0.19


 


     







     

 
  







  



400



p = 0.97
= 0.072






 
  
 


 






 




 






 

400

Amygdala
p = 0.74
= 3.5 





 
 


 

  
 

400

Brain Stem

Caudate
p = 0.54 
= 13

400

Callosum



400

Hippocampus Lat. Ventricle

p = 0.96
p = 0.68


= 0.5
=  210 







 
  
   


 

 


 

 
 



400

Nadir CD4 Count


   




   
  


      
 


  
 



 



 


 
 

400

  





 
 



   



 
  

 
















400

Pallidum
p = 0.54

= 6.3






Putamen
p = 0.88
= 6.2





  
 

 










 



 

  

 






  


 

    

 
   





400







 



 

 
   
 










 
  
 







 

 

400

400

3rd Ventricle



 








  

  


   




   

 

     


 
 


 


 

 
    







 



Thalamus
p = 0.88
= 2.9









p = 0.97
= 0.61




 




  




 




 

 
 












 










     
   


 

  

 
 

p = 0.97

= 0.032



 

p = 0.74
= 36

p = 0.88
= 5.5

 








 


  



   
    


  
 






2
4 p = 0.7

p = 0.91

= 0.66


= 2.3

 





 
  

 






p = 0.88
= 3.7 


 





















30

Amygdala

p = 0.84
= 73























 
 

 

 

  
  










 










   








   

   







   





















30

Brain Stem

30

Caudate

p = 0.9
= 23

30

p = 0.54
= 14

 


  



    
 


 

 
  









 




Callosum
































Hippocampus

p = 0.84


= 210






































30

30

30

Illness Duration (years)

Lat. Ventricle


p = 0.9
= 1000





























Pallidum

p = 0.9

= 20




































 


 



   


 
 







  




 



30

Putamen

p = 0.84
= 210

















































 

 

 
 
 





 
  


    
 
 



p = 0.88
=5








p = 0.84
= 610





 

 
 


  






 





  

 








30 0

Thalamus























30

3rd Ventricle

p = 0.9
= 84

























































p = 0.9

= 40


p = 0.9
= 29






































p = 0.9

= 16













































































p = 0.84
= 160




































p = 0.84

= 150























p = 0.9
= 590






















































p = 0.9
= 7.9























p = 0.9
= 78
















































p = 0.9
= 73




































Viral Load

Combined-Metric Models

p = 0.35
= 9.6

 


4 p = 0.84
= 34

-4
Classification performance
by descriptor












 
  
 




 
 


 



p = 0.54
=  220

-5




 
 




 

   
 
 


   
  
 
 

 
 
 
  


 

30

 

p = 0.54
= 15

p = 0.74
= 8.7




 




 


 
 

  



   

 

  



 



Accumbens

Volume (scaled) mm3










    

  
  




  

 









Thalamus
p = 0.97
= 0.024

Undetectable

Right

t-value

T-value maps of the subcortical surfaces modeling shape differences between HIV+ and control cohorts. First and third
columns illustrate raw t-values from the main effect's coefficient. Second and fourth columns show thresholded surfaces
in which regions not significantly different following FDR correction are depicted in black.




Putamen
p = 0.97
= 0.19

Midline





4 p = 0.9
= 11

Viral Load

Viral Load

-4

Hippocampus Lat. Ventricle

Pallidum
p = 0.97
p = 0.97
p = 0.97



= 0.23
= 7.8
= 0.037

 

HIV+

Left

0
-4

Callosum































Right

t-value

Radial Distance












Detectable

100

Conclusions
Using volumetric and shape-based descriptors we were able to characterize
abnormal subcortical morphometry in HIV.

80

Correctly predicted that more extreme clinical measures were associated with
more extreme subcortical atrophy.

Sensitivity (%)

60

Observed unexpected enlargement of pallidi in relationto certain clinical measures.

40

Volumetric and shape descriptors uniquely characterized separate aspects of the

HIV+ phenotype.
Model

20

Class Prediction

p = 0.044
= 3900

Midline

Caudate
p = 0.97
= 0.38



p = 0.97

=  0.1


= 0.012

Brain Stem





























Left

0
-3

0
-2

AUC

Model

100

80

60

40

20

0 100

Specificity (%)

AUC

Vol. & RD:

55.4
Vol. & JD:
76.44
RD & JD:
68.62
Vol. & RD & JD: 72.87

Vol. AUC: 66.9

RD AUC: 53.79
JD AUC: 74.71

Right

Email: Benjamin.SC.Wade@gmail.com

4 p = 0.97



p(c|v) = 1T pt (c|v)

Single-Metric Models
U

4) Class prediction
based on vote of
terminal nodes




= 0.64

Radial Distance

...

Accumbens

Random Forest Classification

3) Nodes split by
features maximizing
class purity










































 


  





 


    




  
 








2) Sub-tree input
is a bootstrapped
set of observations
and a random
subset of features

v1



















































p = 0.97
= 0.58

 


vt


























Midline

Nadir CD4 Count



























p = 0.089
= 210






















1) RF intput is the
set of all features




































Set of all features, F




 
 

    



     




 


 

 

 

4 p = 0.88
























3rd Ventricle

Left




3) Compute
shape metrics at
each vertex










































Thalamus

p = 0.019
= 390







































31 uninfected controls: 64.68 years old, 2 women

Jacobian Determinant

















































p = 0.89
= 14

Amygdala
p = 0.97
= 0.0082



= 0.11

14

Putamen

p = 0.044
= 330


p = 0.0079
= 420



4 p = 0.97

Subjects
63 elderly HIV+ subjects: 65.35 years old, 4 women

d)

























Pallidum

p = 0.0079

= 180

HIV Status

2) Create mesh
surface for each c)
region







b)

Lat. Ventricle

p = 0.019
= 5100






























p = 0.02
= 290






















1) Segment sub- a)
cortical regions
using FreeSurfer



























































 

Morphometric descriptors

Hippocampus

p = 0.27
= 110





Methods




p = 0.15
= 770


Development of sensitive biomarkers for HIV-related atrophy would aid

clinicians in determining which HIV patients will develop cognitive deficits.





























































Callosum

The extent of impairment is associated with subcortical structural atrophy.

The profile of HIV-associated brain abnormalities remains poorly understood.

Caudate

p = 0.76
= 38


Right

HIV Status










































Brain Stem

Midline

Over 50% of HIV+ individuals show significant neurocognitive impairments.





















p = 0.76
= 19

Left

Can the use of surface-based shape descriptors enhance classification of HIV+

brains beyond the use of volumetric measures?

Introduction

Amygdala

4 p = 0.14
= 41


HIV Status

Are there discernible abnormalities in the subcortical morphometry of elderly

people with HIV?

80

60

40

20

Classification was improved by shape descriptors in some cases.

References

[1] L. Cysique, et al., "Prevalence and pattern of neuropsychological impairment in human immunodeficiency
virus-infected/acquired immunodeficiency syndrome (HIV/AIDS) patients across pre- and post-highly active
antiretroviral therapy eras: A combined study of two cohorts," J. of NeuroVirology, vol. 10, pp. 350-357, 2004.
[2] Y. Wang , et al., "Applying tensor-based morphometry to parametric surfaces can improve MRI-based
disease diagnosis," NeuroImage, vol. 74, pp. 209-230, 2013.
[3] L. G. Apostolova, et al., 3D comparison of hippocampal atrophy in amnestic mild cognitive impairment
and Alzheimer's disease vol. 129, 2006.
Acknowledgements: This work was supported in part by NIH 'Big Data to Knowledge' (BD2K) Center of
Excellence grant U54 EB020403, funded by a cross-NIH consortium including NIBIB and NCI and by the
National Science Foundation Graduate Research Fellowship under Grant No. DGE-0707424 and grant
numbers K23AG032872 (to V.V.), P50 AG023501 (ADRC, PI: Bruce Miller); P30-AI027763 (UCSF CFAR), UL1
RR024131 (UCSF GCRC), the Larry L. Hillblom Foundation and the AIDS Research Institute at UCSF.