Federica Agosta1, Silvia Basaia1, Luca Wagner2, Giuseppe Magnani3, and Massimo Filippi1,3
1Neuroimaging Research Unit, INSPE, Division of Neuroscience, San Raffaele Scientific Institute, Vita-Salute San Raffaele University, Milan, Italy, 2Effeventi, Milan, Italy, 3Department of Neurology, San Raffaele Scientific Institute, Vita-Salute San Raffaele University, Milan, Italy
Synopsis
We built and
validated a deep learning algorithm that predicts the individual diagnosis of Alzheimer’s
disease (AD) and the development of AD in mild cognitive impairment (MCI)
patients based on a single cross-sectional brain structural MRI scan. The deep neural network (DNN) procedure discriminated
AD and heathy controls with an accuracy up to 98%, and MCI converters and MCI
stable with an accuracy up to 75%. DNNs provide a powerful tool for the
automatic classification of AD and MCI prognosis.
Introduction
The utility of structural MRI in the clinical
assessment of patients with suspected Alzheimer’s
disease (AD) will be increased by development of robust algorithms
for automated assessment. Deep learning (DL) algorithms differ from conventional machine
learning methods by the fact that they require little or no image
pre-processing and can automatically infer an optimal representation of the
data from the raw images without requiring prior feature selection, resulting
in a more objective and less bias-prone process. The aim of this study was to build and validate a DL algorithm that predicts the
individual diagnosis of AD and the development of AD in mild cognitive
impairment (MCI) patients based on a single cross-sectional brain structural
MRI scan.Methods
3D
T1-weighted images from ADNI (352 healthy controls [HC], 294 AD, 253 MCI converters,
510 MCI stable) and subjects recruited at our Institute (non-ADNI dataset: 55
HC, 124 AD, 27 MCI converters, 23 MCI stable) were used. Deep neural networks
(DNNs), which are mathematical representations of the human neural architecture
with multiple hidden layers of artificial neurons, were applied. The whole dataset
was randomly divided into a training/validation set (90%) and a testing set
(10%). DNN performance was improved by adding to the original dataset synthetic
images created using data augmentation algorithms, as well as transfer learning
to subsequent comparisons. Results
DNNs with
different architectures and parameters have been optimized and were able to classify with high accuracy HC, AD, MCI
converters, MCI stable. In the reference ADNI dataset and in a second dataset
including also non-ADNI images, the DNN procedure was able to discriminate AD
and HC with an accuracy up to 98%, with no difference between ADNI and non-ADNI
images. In the whole dataset, DNN was also able to discriminate: HC vs (AD + all MCI) with an accuracy up to
86%, and MCI converters vs MCI stable
with an accuracy up to 75%.Discussion
DNN has several
advantages including: ability to process large amount of data, ability to deal
with heterogeneous data, computationally inexpensive (after training) and
reduction of diagnosis time. DNNs are also ideal for real-time applications
where immediate output is desirable.Conclusions
DNNs provide a
powerful tool for the automatic classification of AD patients and MCI
prognosis. Future studies are warranted to test the accuracy of the procedure
in the differential diagnosis between AD and other neurodegenerative diseases.Acknowledgements
The study was supported by the Italian Ministry of Health
(GR-2011-02351217). Data collection and sharing was funded by the Alzheimer’s
Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01
AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012).References
No reference found.