The use of different acquisition protocols induces spurious patterns in the data.
In other words,
different protocols produce a bias in the morphologic measurements.
Using random forest classifiers,
we determined that the influence of the field strength is more significant than that of the MRI-machine manufacturer.
Images obtained using different field strengths can be easily separated with this type of classifier.
Figure 4 shows the 1.5 T and 3 T separation when applying a dimensionality reduction technique called t-SNE [5].
A harmonization algorithm was developed to minimize the influence of the field strength parameter: the z-score calculated for 1.5 T and 3 T for each 10-year-bin was the method that produced the best results in terms of compatibilization.
The effectiveness of the proposed solutions was measured in terms of the degree of confusion achieved during classification.
Results showed that it is necessary to take into account brain measures variability with age.
Best results were achieved when taking into consideration only people under 60 year old,
reaching levels of confounding similar to the estimated chance accuracy.
Figure 5 shows the probabilities of class 3.0 T when using a random forest classifier over the test set,
before (left) and after (right) the compatibilization method using data from people younger than 60 year old.
For this group,
accuracy dropped from 92% to 57%.
The same analysis was conducted with data from people older than 60 year old,
as shown in Figure 6.
In this case,
accuracy did not decrease (below 89%).
The classification rate is substantially reduced after the compatibilization process for people under 60 years old but not for older people.
However,
rather than the accuracy of the classifier,
it is the class probability for the test set that accounts for the effectiveness of the chosen approach.
The probability distribution is narrowed and shifted towards 0.5 which can be thought as a proof for the classifier's increased uncertainty.
A feature importance analysis was conducted in order to determine which variables are not susceptible to the harmonization process.
Indeed,
we found that there is a group of features that are resistant to the compatibilization technique that was proposed.
A preliminary analysis showed that some particular segmentation procedures from FreeSurfer are extremely dependent of the field strength: such as the estimation of blood vessels,
white matter hypointensities.
However these features are not closely linked to a cortical or subcortical structure but to the segmentation algorithms.
Arguably,
the differences in contrast produced by different field strengths affect FreeSurfer segmentation algorithms.
Likewise,
the PFE method demonstrated that some subcortical measurements (fissure of the hippocampus,
thalamus,
ventricles) as well as white matter estimations are important for the class distinction; meaning that their values are the most affected by the differences in the field strength.
The compatibilization process should not affect the power of classifying diseases that alter the brain structure.
Confusion matrices in Figure 7 show the classification results for the data before compatibilization and for the harmonized data.
In order to build the confusion matrix,
we added up the test outputs for each fold.
Apart from slightly reducing the number of false positives,
applying the compatibilization methods does not affect the classification of pathologies.
The feature importance analysis showed that the variables that were useful for classification were those with a well-known clinical relevance for AD,
namely: hippocampal volumes as well as volumes and thickness from fusiform,
inferior temporal and entorhinal cortex.