Keywords:
Artificial Intelligence, Anatomy, Neuroradiology brain, MR, Segmentation, Computer Applications-General, Computer Applications-Detection, diagnosis, Dementia
Authors:
I. Evangelista, C. L. Galimberti, G. Pascariello, J. C. Gomez, A. L. Rodríguez Musso, P. Donnelly-Kehoe; Rosario/AR
DOI:
10.26044/ecr2019/C-2108
Methods and materials
FreeSurfer [1] is a software package that can enhance the visual information in resonance images by means of performing structural measurements.
Using FreeSurfer v6.0,
2700 MRIs were processed and 272 morphometric features that characterize each brain's cortical and subcortical anatomy were obtained.
An example of FreeSurfer segmentation is shown in Figure 2,
where each colour represents a cortical or subcortical structure.
From the segmentation of these structures,
thickness,
area and volume values can be measured.
The effects of the different acquisition parameters on the morphological variables were tested using dimensionality reduction approaches and supervised machine learning techniques (random forests).
We trained random forest classifiers implemented in Python's Scikit-Learn [2] using the neuromophometric measurements of HC as features.
The classification objective was to determine certain acquisition parameter,
for example the field strength (1.5T vs 3T) or manufacturer (Simens,
General Electric or Philips); that is to say,
a classifier was supposed to determine an acquisition parameter based solely on structural information acquired using FreeSurfer.
Classification accuracy was taken as an indicator of the extent to which brain measures are different between classes.
Since a random forest classifier outputs classification probabilities,
these values were also taken as an indicator of the degree of compatibilization achieved.
In simpler words,
the more certain the classifier was about the output class,
the stronger the differences between the classes.
Different solutions were explored to achieve compatibilization: z-score using healthy controls of each group to determine the mean and standard deviation; w-score using healthy controls of each group to determine the linear regression; z-score taking into account age (greater or lower than certain threshold).
Moreover,
given the natural aging process of the brain,
it was considered essential that the compatibilization technique took into account the changes that brain structures experience with age [3].
We ultimately proposed binning the age and calculating a z-score for each group using healthy controls to determine the mean and standard deviation.
Figure 3 shows how the morphometric values change with age for a randomly chosen structure (medial orbitofrontal cortex),
and the means calculated for 1.5 T and 3 T for each 10-year bin.
Further,
we performed a progressive feature elimination (PFE) [4] in order to determine which variables were important for distinguishing data acquired with different parameters and to analyse the presence of features that were invulnerable to the compatibilization process,
i.e.
whether there were features that contained information for separating the classes in spite of the proposed compatibilization method.
Finally,
in order to corroborate that the data classification power was not modified after the compatibilization,
we analyzed how the data was classified depending on the diagnosis: HC,
MCI and AZ.
Random forest classifiers were developed using one third of the data to perform feature selection and a 10-fold cross validation to evaluate the model.