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
Aims and objectives
Structural MRI (sMRI) of the brain in the clinical environment usually consist in a qualitative analysis,
in which the neuroradiologist detects deviations or anomalies by visual inspection.
A quantitative analysis can complement the specialist's opinion,
allowing a comparative study among subjects.
However,
to perform a reliable analysis,
it is necessary to consider each measure’s distribution in the healthy population.
In this work we used a large sample from several public databases to study the influence of acquisition parameters on the measurement of neuromorphometric features computed with Freesurfer.
Then,
we applied them to study Alzheimer's disease (AD) and Mild Cognitive Impairment (MCI) from a machine learning perspective.
Quantitative approaches to sMRI are ideal for studying changes in brain structures due to neurodegenerative diseases.
For the purpose of developing a system that can help to assess structural abnormalities for individualized medicine,
2700 T1-MRI images were collected from public databases.
These data correspond to healthy controls (HC) and both AD and MCI patients.
The T1-MRI were acquired under different protocols according to the procedures of the center where participants were scanned.
Figure 1 shows the age distribution of participants included in this analysis.
From each T1-MRI a set of quantitative features characterising each brain’s anatomy was obtained using FreeSurfer v6.0.
While open databases allow to collect a vast amount of images,
combining data from different sources is not straightforward.
Different acquisition parameters may bias Freesurfer’s morphologic measurements [1].
Indeed,
T1-MRI brain images obtained using different field strengths were classified with 88% of accuracy.
According to our analysis,
differences introduced by field strength were more relevant than those caused by other factors,
including the manufacturer,
the machine model and the voxel volume.
The possibility of having images obtained using different acquisition parameters is important in the interest of achieving generalization.
However,
in order to attain a system that can prove robust against changes in acquisition parameters it is required to develop a technique that can minimize the differences caused by them.
Therefore,
different strategies were explored with the objective of harmonizing data from different sources.
The effectiveness of the proposed solutions was analysed from a machine learning perspective.
Results showed that it is necessary to take into account brain measures variability with age so as to properly mitigate the influence of the acquisition parameters.
Harmonization between field strength groups makes it easier to compare images from different sources,
albeit there are a number of variables that cannot be used indistinctly.
Thus,
an analysis of feature importance was conducted in order to determine which variables are not susceptible to the compatibilization process.
What is expected from a harmonization process is that it does not affect the the capacity of classifying diseases using the data.
Morphological measurements after applying the compatibilization technique proved as effective as the original data to classify AD and MCI.