【论文阅读】Automated detection of Alzheimer disease using MRI images and deep neural networks- A review

Early detection of Alzheimer’s disease is crucial for deploying interventions and slowing the disease progression. The purpose of this survey is to provide an overview of recent research on deep learning models for Alzheimer’s disease diagnosis.

The first step in the automated detection of Alzheimer’s is feature selection–choosing the right features to feed the deep learning model. MRI images are often 3D, and thus result in large feature space, making feature selection an essential component. The traditional methods of extracting features are frequently constrained by some a priori information, which means that they can only extract certain aspects tha relate to a particular application.

Using MRI, it is also possible to observe the disease’s effect on the subject’s spatial domain and temporal region, which create patterns.
The degeration in the hippocampal region is easily identifiable in Alzheimer’s disease patients compared to non-AD patients.

技术层面：
The models are categorized as CNN based, using transfer learning, hybrid models, and other deep learning and machine learning techniques.
CNN模型发展从3DCNN、MSCNet(multi-scale CNN)、3D ShuffleNet、AlzNet、DEMentia NETwork(VGG16+Ensemble based classifier+SiameseNetwork)
Hybrid Models 发展 从PCANet+kMeans、3DPCANet、3DPCANet+SVM+mALFF、Sparse Autoencoder+Genetic Algorithm+SVM
其他有意思的算法，Hierarchal Extreme Machine Learning, minimum-redundancy maximum-relevance(mRMR)+ReliefF algorithm+SVM, Behavior score-embedded encoder network, Brain functional connectivity network(BFCN) + attention-based bidirectional LSTM, BrainNet2D, Deep cross-model attention network(DCMAT), Ensemble DCNNs of multi-modality images by Adaboost learner, GCN-EMCI

相关可参考的代码

结论：Recent trend in research of using end to end deep learning methods present great potential in the early detection of MCI.
Challenging: compare different review papers because of highly variable methodologies including different datasets, preprocessing methods, inputs selected and also the final task of classification or segmentation.
Futures: spatio-temporal maps, structural brain network, multimodal models.

个人感受，整篇文章写得最有感触的是结论部分里面的challenging，实在是太难compare了