Please Wait...

Assessing Regional Cortical Thickness for Predicting MCI Conversion to AD

Joel Schaerer, Mehul Sampat, Florent Roche, Joonmi Oh, Luc Bracoud, Joyce Suhy and the Alzheimer's Disease Neuroimaging Initiative

Bioclinica, Newark, CA, USA and Lyon, France

 

BACKGROUND

Brain cortical thickness is gaining popularity as an endpoint in AD clinical trials. While a global approach provides valuable insights, a regional analysis may be more sensitive. The aim of this work is to determine if a regional approach is preferable and potentially which aggregation of regions is the most sensitive for prediction of conversion to AD.

METHODS

Cortical thickness measurements were performed on the Baseline 3DT1-weighted sequences of 232 MCI subjects (116 who did not convert to AD within 36 months, 116 who did) from the ADNI-1 database (http://adni.loni.ucla.edu).


Image processing
FreeSurfer v5.3 (http://surfer.nmr.mgh.harvard.edu/) [1, 2] was used for the determination of the white/gray boundary (in blue below), pial/CSF boundary (in green below) and subcortical parcellation (see Fig. 1). Cortical thickness was assessed by solving Laplace’s equation [3], building the set of paths between the inner and outer cortical surfaces and deriving the average thickness from them, for each subcortical region provided by FreeSurfer, as well as overall.

Statistical analysis
Performance of the region(s) of interest in discriminating between converters and non-converters was tested via ROC analyses using the Area Under the ROC Curve (AUC).

The Lasso method [4] was used in order to optimally combine regions. This consists in optimally weighting each region in order to best predict conversion to AD using a composite cortical thickness score. The function to minimize consists of a data term reflecting prediction accuracy which aims at rewarding good predictivity, and a model term which aims at controlling for the number of regions used and potentially penalizing models with too many regions. This model term is controlled by a regularization coefficient λ. A leave-one-out strategy was used in order to estimate the optimal λ value.

Using this optimal λ, the model was fitted over the whole database for determination of the corresponding list of regions and coefficients, thus defining an optimal model.

A threefold cross-validation strategy run 100 times with the corresponding λ was used in order to estimate the AUC for this model.

RESULTS

First looking at the average cortical thickness over the whole cortex, analyses of ROC curves yielded an AUC of 0.664 to distinguish between MCI converters and non-converters.
Moving to a more regional analysis, we then looked at each region individually. Of the available regions from the FreeSurfer atlas, the Right Inferior Temporal was the most sensitive (AUC = 0.698). Results for the next best other individual regions are shown in Table 1.

In order to look for the most optimal combination of regions for the prediction of conversion, the Lasso method determined an optimal λ value (see Figure 2). When applied to the whole database, this led to a composite region made of 15 subregions (see Table 2).

Cross-validation estimated the AUC for this model at 0.700, which was only marginally higher in sensitivity than the best individual region.

CONCLUSION

While a regional analysis of cortical thickness best predicts conversion to AD, compared to whole cortex analysis, looking for an optimal aggregation did not provide significantly better results. Nevertheless, such approach may be more robust to image quality issues, which is to be further confirmed on other datasets.
This methodology could be used for other applications, such as discriminating different patient subgroups using other endpoints. In particular, a similar analysis will be performed on longitudinal data in order to determine which single or composite region would provide the best effect size when measuring change in cortical thickness over time.

 

REFERENCES

[1] Fischl et al., Whole Brain Segmentation: Automated Labeling of Neuroanatomical Structures in the Human Brain, Neuron 2002

[2] Fischl et al., Automatically Parcellating the Human Cerebral Cortex, Cerebral Cortex 2004

[3] Jones et al., Three-Dimensional Mapping of Cortical Thickness Using Laplaces Equation, Human Brain Mapping 2000.

[4] Tibshirani et al., Regression Shrinkage and Selection via the Lasso, J. Royal. Statist. Soc B. 1996

LEARN MORE OR SPEAK WITH OUR EXPERTS

CONTACT US
Leader in Clinical Trial
Management Solutions

Successful clinical trials require the ability to see key details and uncover hidden insights. Bioclinica utilizes science and technology to bring clarity to clinical trials, helping companies to develop new life-improving therapies more efficiently and safely.

ASCPT 2018 Annual Meeting https://t.co/i0mjYbBq5N Join @Bioclinica Research's Open House 3/22 https://t.co/citlLotbD7
bioclinica (2 days ago)
Heading to DC for DIA's #Pharmacovigilance & Risk Mgmt Strategies Conf.? Visit @Bioclinica in booth 32. Preview our… https://t.co/alU1sCg04l
bioclinica (2 days ago)
Kevin Bishop, named a 'Top Innovator' in #clinicaltrials by @CenterWatch See how we're reducing risk to the clinic… https://t.co/rXBGvIh9Wr
bioclinica (3 days ago)
RT @CenterWatch: Congratulations to all our #2018TopInovators! Read about the innovations in the #CWMonthly special edition, https://t.co/Y…
bioclinica (2 weeks ago)
RT @CenterWatch: .@Pfizer demonstrates leadership in #clinicaltrial innovation with two spots on the @CenterWatch #2018TopInnovators, https…
bioclinica (2 weeks ago)
RT @jheanne: I just contributed towards 'Joby' on https://t.co/EHt69JFLZ9 Best wishes to Joby and all our @Bioclinica team members! @Oxfam…
bioclinica (2 weeks ago)

Latest Blogs:

Removing Risk from Clinical Trial Management System (CTMS) Implementations
Collaboration Between Clinical Operations and the Logistics and Supply Chain Teams is Key to Trial Success
The Value of Protocol Review
CTMS and RBM: Hot Topics at OCT Nordics in Copenhagen