Camilla Cividini1, Federica Agosta1, Silvia Basaia1, Luca Wagner1, Maura Cosseddu2, Elisa Canu1, Stefano Gazzina2, Giuseppe Magnani3, Elka Stefanova4, Vladimir S. Kostic4, Roberto Gasparotti5, Alessandro Padovani2, Barbara Borroni2, and Massimo Filippi1,3
1Neuroimaging Research Unit, Institute of Experimental Neurology, Division of Neuroscience, San Raffaele Scientific Institute, Vita-Salute San Raffaele University, Milan, Italy, 2Neurology Unit, Department of Clinical and Experimental Sciences, University of Brescia, Brescia, Italy, 3Department of Neurology, Institute of Experimental Neurology, Division of Neuroscience, San Raffaele Scientific Institute, Vita-Salute San Raffaele University, Milan, Italy, 4Clinic of Neurology, Faculty of Medicine, University of Belgrade, Belgrade, Yugoslavia, 5Neuroradiology Unit, Department of Clinical and Experimental Sciences, University of Brescia, Brescia, Italy
Synopsis
We built and validated a convolutional neural network (CNN) to predict
the individual diagnosis of early-onset Alzheimer’s disease (EOAD) and
behavioral variant of frontotemporal dementia (bvFTD) based on a single
T1-weighted image. The analysis showed that CNN procedure was able to
discriminate EOAD from healthy controls with an accuracy of 83%
(sensitivity=85% and specificity=80%). CNNs differentiated bvFTD patients from
controls with an accuracy of 73% (sensitivity=63% and specificity=83%). CNNs
provide a powerful tool for the automatic classification of early-onset
neurodegenerative dementia and perform well without any prior feature
engineering and regardless the variability of imaging protocols and scanners.
Introduction
Convolutional neural networks (CNNs) are mathematical representations of
the human neural architecture with multiple hidden layers of artificial
neurons, mimicking how the human brain processes information. Their use achieved
optimal results in many domains, such as speech recognition tasks, computer
vision and, more recently, diagnostic systems and biomedical imaging. In this
study, we built and validated a deep learning algorithm that predicts the
individual diagnosis of early-onset neurodegenerative dementia based on a
single 3D T1-weighted image.Methods
The study involved patients
with early-onset Alzheimer’s disease (EOAD) and behavioral variant of frontotemporal
dementia (bvFTD) and healthy controls. 3D T1-weighted images were obtained from
ADNI (75 EOAD and 361 controls) and subjects recruited from three non-ADNI centers
(80 EOAD, 52 bvFTD, and 139 controls). Overall, the sample included 155 EOAD
patients, 52 bvFTD patients and 500 controls. CNNs were applied on 3D
T1-weighted images. 3DT1-weighted images were normalized to the MNI space using
Statistical Parametric Mapping (SPM12) and the Diffeomorphic Anatomical
Registration Exponentiated Lie Algebra (DARTEL) registration method. T1 images
were segmented to produce gray matter, white matter and cerebrospinal fluid
tissue probability maps in the MNI space. The segmentation parameters were
imported in DARTEL and the rigidly aligned version of the images was generated;
the DARTEL template was created and the obtained flow fields were applied to
the modulated 3DT1-weighted images of single subjects to warp them to the
common DARTEL space. Subsequently, images from DARTEL were normalized to the
MNI template using an affine transformation estimated from the DARTEL gray
matter template and the a priori gray matter probability map. The CNNs
implemented in this study had a network architecture that used 3D convolutions
because of the volumetric nature of MRI images. The inputs were the normalized
3DT1-weighted images and the outputs to be predicted were the subject groups. The
architecture of the network contained: 12 repeated blocks of convolutional
layers, 2 blocks with 50 kernel of size 5x5x5 and 10 block with 100 to 1600
kernels of size 3x3x3, a Rectified Linear Unit as activation layer, a fully
connected layers and one output layer designed as logistic regression. All
software was written in Python using Theano, a scientific computing library
with support for machine learning and GPU computing. The dataset was randomly
divided into training/validation set (80%) and testing set (20%). CNN performance
was improved by adding to the original dataset synthetic images created using
data augmentation algorithms. CNN performance was evaluated by sensitivity,
specificity and accuracy.Results
CNNs with different architectures and parameters were optimized. The analysis showed that CNN procedure
was able to discriminate EOAD from healthy controls with an accuracy of 83%
(sensitivity=85% and specificity=80%). CNNs differentiated bvFTD patients from controls
with an accuracy of 73% (sensitivity=63% and specificity=83%).Discussion
Our CNN was highly-performing in differentiating EOAD and bvFTD from
healthy controls. Importantly, the algorithm performed well without any prior
feature engineering and regardless the variability of imaging protocols and
scanners.Conclusions
CNNs have several advantages,
including the ability to process large amount of data and to reduce diagnosis
time. Furthermore, the algorithm demonstrated the capability to be
exploitable by not-trained operators and likely to be generalizable to unseen
patient data. CNNs provide a
powerful tool for the automatic classification of early-onset neurodegenerative
dementia. Future studies are warranted to test the accuracy of the procedure in
differentiating among diseases.Acknowledgements
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). The study was supported by the Italian Ministry
of Health (GR-2011-02351217; GR-2013-02357415).References
No reference found.