Maybe there is something here but I don't know. This sentence leads to this research but I see nothing here I can understand.
An artificial intelligence model accurately predicted Alzheimer's risk by assessing MRIs, cognitive scores, age, and sex, besting 11 practicing neurologists.
Development and validation of an interpretable deep learning framework for Alzheimer’s disease classification
Brain, awaa137, https://doi.org/10.1093/brain/awaa137
Published:
01 May 2020
Introduction
Millions worldwide continue to suffer from Alzheimer’s disease (Scheltens et al., 2016),
while attempts to develop effective disease-modifying treatments remain
stalled. Though tremendous progress has been made towards detecting
Alzheimer’s disease pathology using CSF biomarkers (Frisoni et al., 2010; Jack et al., 2013; Harper et al., 2014), as well as PET amyloid (Nordberg, 2004; Bohnen et al., 2012), and tau imaging (Mattsson et al., 2019; Ossenkoppele et al., 2019),
these modalities often remain limited to research contexts. Instead,
current standards of diagnosis depend on highly skilled neurologists to
conduct an examination that includes inquiry of patient history, an
objective cognitive assessment such as bedside Mini-Mental State
Examination (MMSE) or neuropsychological testing (McKhann et al., 2011), and a structural MRI to rule in findings suggestive of Alzheimer’s disease (Frisoni et al., 2010).
Clinicopathological studies suggest the diagnostic sensitivity of
clinicians ranges between 70.9% and 87.3% and specificity between 44.3%
and 70.8% (Beach et al., 2012).
While MRIs reveal characteristic cerebral changes noted in Alzheimer’s
disease such as hippocampal and parietal lobe atrophy (Whitwell et al., 2012), these characteristics are considered to lack specificity for imaging-based Alzheimer’s disease diagnosis (van de Pol et al., 2006; Barkhof et al., 2007; Raji et al., 2009; Frisoni et al., 2010).
Given this relatively imprecise diagnostic landscape, as well as the
invasive nature of CSF and PET diagnostics and a paucity of clinicians
with sufficient Alzheimer’s disease diagnostic expertise, advanced
machine learning paradigms such as deep learning (LeCun et al., 2015; Hinton, 2018; Topol, 2019), offer ways to derive high accuracy predictions from MRI data collected within the bounds of neurology practice.
Recent
studies have demonstrated the application of deep learning approaches
such as convolutional neural networks (CNNs) for MRI and multimodal
data-based classification of cognitive status (Qiu et al., 2018).
Despite the promising results, these models have yet to achieve full
integration into clinical practice for several reasons. First, there is a
lack of external validation of deep learning algorithms since most
models are trained and tested on a single cohort. Second, there is a
growing notion in the biomedical community that deep learning models are
‘black-box’ algorithms (Castelvecchi, 2016).
In other words, although deep learning models demonstrate high accuracy
classification across a broad spectrum of disease, they neither
elucidate the underlying diagnostic decisions nor indicate the input
features associated with the output predictions. Lastly, given the
uncertain onset and heterogeneity of symptoms seen in Alzheimer’s
disease, a computerized individual-level characterization of Alzheimer’s
disease remains unresolved. Considering these factors, we surmise that
the clinical potential of deep learning is attenuated by a lack of
external validation of single cohort-driven models, and an increasing
use of opaque decision-making frameworks. Thus, overcoming these
challenges is not only crucial to harness the potential of deep learning
algorithms to improve patient care, but to also pave the way for
explainable evidence-based machine learning in the medical imaging
community. To address these limitations, we developed a novel deep
learning framework that links a fully convolutional network (FCN) to a
traditional multilayer perceptron (MLP) to generate high-resolution
visualizations of Alzheimer’s disease risk that can then be used for
accurate predictions of Alzheimer’s disease status (Fig. 1).
Four distinct datasets were chosen for model development and
validation: Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset,
Australian Imaging, Biomarker and Lifestyle Flagship Study of Ageing
(AIBL), Framingham Heart Study (FHS), and National Alzheimer’s
Coordinating Center (NACC) (Table 1 and Supplementary Fig. 1).
Association of model predictions with neuropathological findings along
with a head-to-head comparison of the model performance with a team of
neurologists underscored the validity of the deep learning framework.
Figure 1
Schematic of the deep learning framework.
The FCN model was developed using a patch-based strategy in which
randomly selected samples (sub-volumes of size 47 × 47 × 47 voxels) of T1-weighted
full MRI volumes were passed to the model for training (Step 1). The
corresponding Alzheimer’s disease status of the individual served as the
output for the classification model. Given that the operation of FCNs
is independent of input data size, the model led to the generation of
participant-specific disease probability maps of the brain (Step 2).
Selected voxels of high-risk from the disease probability maps were then
passed to the MLP for binary classification of disease status (Model A
in Step 3; MRI model). As a further control, we used only the
non-imaging features including age, gender and MMSE and developed an MLP
model to classify individuals with Alzheimer’s disease and the ones
with normal cognition (Model B in Step 3; non-imaging model). We also
developed another model that integrated multimodal input data including
the selected voxels of high-risk disease probability maps alongside age,
gender and MMSE score to perform binary classification of Alzheimer’s
disease status (Model C in Step 3; Fusion model). AD = Alzheimer’s
disease; NC = normal cognition.
Table 1
Study population and characteristics
| Dataset | ADNI | AIBL | FHS | NACC | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Characteristic | NC | AD | P-value | NC | AD | P-value | NC | AD | P-value | NC | AD | P-value |
| (n = 229) | (n = 188) | (n = 320) | (n = 62) | (n = 73) | (n = 29) | (n = 356) | (n = 209) | |||||
| Age, years, median [range] | 76 [60, 90] | 76 [55, 91] | 0.4185 | 72 [60, 92] | 73 [55, 93] | 0.5395 | 73 [57, 100] | 81 [67, 94] | <0.0001 | 74 [56, 94] | 77 [55, 95] | 0.0332 |
| Education, years, median [range] | 16 [6, 20] | 16 [4, 20] | <0.0001 | NAa | NAa | NA | 14 [8, 25] | 13 [5, 25] | 0.3835 | 16b [0, 22] | 14.5b [2, 24] | 0.8363 |
| Gender, male (%) | 119 (51.96) | 101 (53.72) | 0.7677 | 144 (45.00) | 24 (38.71) | 0.40 | 37 (50.68) | 12 (41.38) | 0.51 | 126 (35.39) | 95 (45.45) | 0.0203 |
| MMSE, median [range] | 29 [25, 30] | 23.5 [18, 28] | <0.0001 | 29 [25, 30] | 21 [6, 28] | <0.0001 | 29c [22, 30] | 25c [10, 29] | <0.0001 | 29c [20, 30] | 22c [0, 30] | <0.0001 |
| APOE4, positive (%) | 61 (26.65) | 124 (65.97) | <0.0001 | 11 (3.44) | 12 (19.35) | <0.0001 | 13 (17.81) | 11d (40.74) | 0.035 | 102 (28.65) | 112 (53.59) | <0.0001 |
Four
independent datasets were used for this study including: the ADNI
dataset, the AIBL, the FHS, and the NACC. The ADNI dataset was randomly
split in the ratio of 3:1:1, where 60% of it was used for model
training, 20% of the data were used for internal validation and the rest
was used for internal testing. The best performing model on the
validation dataset was selected for making predictions on the ADNI test
data as well as on the AIBL, FHS and NACC datasets, which served as
external test datasets for model validation. All the MRI scans
considered for this study were performed on individuals within ±6 months
from the date of clinical diagnosis. AD = Alzheimer’s disease; NA = not
available; NC = normal cognition.
No comments:
Post a Comment