A Practical Alzheimer Disease Classifier via Brain Imaging-Based Deep Learning on 85,721 Samples

CAS Key Laboratory of Behavioral Science, Institute of Psychology, Beijing, China; Department of Psychology, University of Chinese Academy of Sciences, Beijing, China; Center for Cognitive Science of Language, Beijing Language and Culture University, Beijing, China; Sino-Danish College, University of Chinese Academy of Science, Beijing, China; Sino-Danish Center for Education and Research, Beijing, China; Department of Neurosurgery, Huashan Hospital, Fudan University, Shanghai Neurosurgical Clinical Center, Shanghai, China; Department of Radiology, The First Affiliated Hospital, College of Medicine, Zhejiang University, Hangzhou, Zhejiang, China; Imaging

We further investigated whether the AD classifier could predict disease progression in people 2 4 0 with MCI. MCI is a diagnosis defined as cognitive decline without impairment in everyday 2 4 1 activities 15 . People with the amnestic subtype of MCI have a high risk of converting to AD.

4 2
We screened imaging records of the MCI patients who converted to AD later in the ADNI and images labeled as 'AD' after conversion were not included). We also assembled 4,018 2 4 6 samples from 524 participants labeled 'sMCI' without later progression for contrast. We directly fed all these MCI images into the AD classifier without further fine-tuning, thus 2 4 8 evaluating the performance of the AD classifier on unseen MCI information. To better understand the brain imaging-based deep learning classifier, we calculated 2 5 2 occlusion maps for the classifiers. We repeatedly tested the images in testing sample using 2 5 3 the model with the highest accuracy within the 5 folds, while successively masking brain 2 5 4 areas (volume = 18mm*18mm*18mm, step = 9mm) of all input images. The accuracy 2 5 5 achieved on "intact" samples by the classifier minus accuracy achieved on "defective" 2 5 6 samples indicated the "importance" of the occluded brain area for the classifier. The coefficient between the predicted scores and mini-mental state examination (MMSE) scores 2 6 0 of AD, NC and MCI samples. We also used general linear models (GLM) to verify whether 2 6 1 the predicted scores (or MMSE score) showed a group difference between people with sMCI 2 6 2 and pMCI. The age and sex information of MCI participants was included in this GLM as 2 6 3 covariates. We selected the T1-weighted images from the first visit for each MCI subject and 2 6 4 12 finally collected data from 243 pMCI patients and 524 sMCI patients. Only brain imaging data with enough size and variety can make deep learning accurate and 2 6 9 robust enough to build a practical classifier. We received permissions from the administrators  we did not feed raw data into the classifier for training, but used prior knowledge regarding 2 8 3 helpful analytic pipelines. The brain structural data were segmented and normalized to yield 2 8 4 grey matter density (GMD) and grey matter volume (GMV) maps (i.e., GMD and GMV under the curve (AUC) of the receiver operating characteristic (ROC) curve reached 0.981 2 9 0 (see Fig. 2). In short, our model can classify the sex of a participant based on brain structural 2 9 1 imaging data from anyone and any scanner with an accuracy of about 95%. Interested readers 2 9 2 can test this model on our online prediction website (http://brainimagenet.org). of the sex classifier in the validation sample. After creating a practical brain imaging-based classifier for sex with high cross-dataset samples were taken together (see Fig. 3 and Table 1 and calculated the accuracy difference between the two groups for 100,000 times, see  in the validation sample. To test the generalizability of the AD classifier, we applied it to unseen independent AD 3 2 0 datasets, i.e., AIBL and OASIS 1 and 2. The AD classifier achieved 94.2% accuracy in AIBL 3 2 1 16 with 0.97 AUC (see Fig. 4A). Sensitivity and specificity were 0.881 and 0.954, respectively.

2 3
Sensitivity and specificity were 0.796 and 0.902, respectively.  Importantly, although the AD classifier is agnostic to brain imaging data of MCI, we directly patients who were predicted as AD was considered as sensitivity and the percentage of sMCI 3 3 6 patients who were predicted as AD was considered as 1-specificity, the AUC of ROC curve 3 3 7 for AD classifier reached 0.82. These results suggest that the classifier is practical for screening MCI patients who have a higher risk of progression to AD. In sum, we believe our  Using an unprecedentedly diverse brain imaging sample, we pre-trained an industrial-grade classifier also showed the potential to predict the prognosis of MCI patients. The industrial-grade high accuracy and generalization capability of our deep neural network 4 1 6 classifiers demonstrate that brain imaging did have practical utility for auxiliary diagnosis. The current prototype may facilitate future research to apply brain imaging in many practical 4 1 8 application fields. Of note, the output of the deep neural network model is a continuous 4 1 9 variable, so the threshold can be adjusted to balance sensitivity and specificity. For example, the false-negative rate should be minimized even at the cost of higher false-positive rates. If we lower the threshold (e.g., to 0.2), sensitivity can be improved to 0.881 at a cost of was able to quantify key disease milestones by predicting disease progression in MCI patients.

3 3
In fact, people with pMCI were 3 times more likely to be classified as AD than sMCI ( Although deep-learning algorithms have often been described as "black boxes" for their poor 4 5 4 interpretability, our subsequent analyses showed that the current MRI-based AD biomarker improved performance of the optimized AD classifier (see Fig. 5 and Fig. S4 to compare the 4 6 5 occlusion maps). performance (see Fig. S2-3). There is some reported evidence that truncating or pruning 4 7 2 models before transfer learning may facilitate the performance of the transferred models 29,30 .

7 3
As the sample for training the AD classifier is considerably smaller than that used to train the By precisely predicting the sex of people, the present study also advances our understanding features from structural MRI and concluded that "the so-called male/female brain" does not However, human brains may embody sexually dimorphic features in a multivariate manner. shows that the "male/female brain" does exist, in the sense that accurate classification is In the deep learning field, the appearance of ImageNet tremendously accelerated the study confirms that the "pre-train + fine-tuning" paradigm does work for MRI-based 4 9 6 auxiliary diagnosis. Unfortunately, no such well-preprocessed dataset exists in brain imaging 4 9 7 domain. As data organization and preprocessing of MRI data require tremendous time, 4 9 8 manpower and computational load, these constraints impede scientists from other fields 4 9 9 entering brain imaging. Open access to large amounts of preprocessed brain imaging data is and sharing a practical brain imaging-based deep learning classifier, we would openly share 5 0 2 all sharable preprocessed data to invite researchers (especially computer scientists) to join the 5 0 3 efforts to create predictive models using brain images (Link_To_Be_Added upon publication, 5 0 4 preprocessed data of some datasets could not be shared as the raw data owners do not allow 5 0 5 sharing of data derivatives). We anticipate that this dataset may boost the clinical utility of 5 0 6 brain imaging as ImageNet has done in computer vision research. We openly share our work. Finally, we have also built a demonstration website for classifying sex and AD  P  r  e  d  i  c  t  i  n  g  t  h  e  p  r  o  g  r  e  s  s  i  o  n  o  f  m  i  l  d  c  o  g  n  i  t  i  v  e  i  m  p  a  i  r  m  e  n  t  u  s  i  n  g  m  a  c  h  i  n  e  l  e  a  r  n  i  n  g  :  A   6  5  3   s  y  s  t  e  m  a  t  i  c  ,  q  u  a  n  t  i  t  a  t  i  v  e  a  n  d  c  r  i  t  i  c  a  l  r  e  v  i  e