Research Yan Chaogan's research group developed a practical Alzheimer's disease brain imaging deep learning discriminator

Recently, Yan Chaogan's research group of the Key Laboratory of Behavioral Sciences, Chinese Academy of Sciences, published a paper
entitled "Building a practical Alzheimer's disease brain imaging deep learning discriminator based on 85,721 samples" in the Journal of Big Data (IF=10.
835), a top journal in the field of data science.

Magnetic Resonance Imaging (MRI) is a non-invasive, radiation-free imaging technique that is widely used in the clinical diagnosis of
brain injuries and brain tumors.
However, for brain diseases such as Alzheimer's disease (AD), which cannot be directly distinguished by the human eye, there has been little
progress in MRI-based auxiliary diagnosis.
This is because the vast majority of AD classification algorithms use less training data, and the data tends to come from a few sites, which means that the generalization and accuracy of the classifier will be greatly reduced
when applied to other unfamiliar scanners/populations.
In order to solve this problem, Yan Chaogan's research group based on the transfer learning framework uses a magnetic resonance brain image dataset with unprecedented sample size and diversity to establish a practical AD classifier
based on magnetic resonance structure images.

Constructing a deep learning model with good generalization requires a large amount of training data
.
Yan Chaogan's research group collected 85,721 MRI scans from 50,876 participants from more than 217 sites/scanners by applying for a public dataset, which is one of
the largest sample studies in the field of brain imaging.
Subsequently, the research group used deep convolutional neural networks to construct a gender classifier based on gray matter thickness and gray matter volume index maps, and proposed the concept of
"brain gender".
The gender classifier achieves 94.
9% accuracy in cross-validation across datasets, that is, the model can predict the sex
of any participant from any scanner at any site with 94.
9% accuracy with gray matter volume and gray matter density maps.

Then, using the gender classifier as the benchmark model, the researchers constructed the AD classifier on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset based on transfer learning, so that it achieved 90.
9% accuracy in cross-site cross-validation and 94.
5%, 93.
6%, and 91.
1% accuracy on three independent datasets, respectively , showing strong practicality
.
Using this classifier directly to test structural imaging samples from patients with mild cognitive impairment (MCI), MCI patients who progressed to AD were more than 3 times more likely to be predicted to have AD (65.
2%), more than 3 times that of patients with unprogressed MCI (20.
6%).

In addition, the predictive score of the AD classifier shows a significant correlation with disease severity and has the potential
to be practical in clinical screening.
The relevant sex classifier and AD classifier have been deployed to an online test website http://brainimagenet.
org, interested researchers only need to upload raw data or preprocessed gray matter volume and gray matter density map to obtain the corresponding brain sex or AD discriminant score (at this stage, the discriminant score is only used for scientific research purposes, not as clinical diagnosis).

。 Research code and models have been open-sourced shared in https://github.
com/Chaogan-Yan/BrainImageNet, preprocessing data has been shared in http://rfmri.
org/BrainImageNetData, and will soon be uploaded to the Psychological Science Data Bank
。

In the future, Yan Chaogan's research group will continue to explore the application of the concept of "brain gender" to gender groups and brain disease groups, and further improve the ability of
AD classifiers to predict disease progression.
The classifier is expected to be able to make non-invasive and non-radiation magnetic resonance images with high penetration rate assist or replace invasive radiation positron emission tomography (PET) examination, improve the early diagnosis efficiency of AD, and create greater social value
.