Source-linked AI summary
Deep Learning in Alzheimer's disease: Diagnostic Classification and Prognostic Prediction using Neuroimaging Data
Taeho Jo, Kwangsik Nho, Andrew J. Saykin
TL;DR
Early AD detection and classification require approaches that can handle complex neuroimaging data while reducing the specialized, multi-stage preprocessing burden of traditional machine learning. This systematic review analyzed 16 studies published from 2013 to 2018, finding that deep learning and hybrid approaches achieved strong classification and conversion-prediction performance, while transparency and reproducibility remained challenges.
Problem
Traditional machine-learning classification requires specialized knowledge and multiple stages of feature extraction, selection, dimensionality reduction, and optimization for complex neuroimaging data.
Method
The review analyzed 16 publications and classified their deep-learning approaches according to algorithms and neuroimaging types, comparing binary classification results and related study characteristics.
Results
The review identified 4 hybrid studies combining deep learning with traditional machine learning and 12 studies using only deep learning approaches.
Takeaways & Limitations
Deep learning can identify diagnostic features directly from input images without human intervention, supporting its increasing use in medical image analysis.
Takeaways & Limitations
Deep-learning models remain difficult to interpret, and limited neuroimaging data requires architectural care to avoid overfitting, data leakage, and reproducibility problems.
Abstract
from arXiv · showhide
Deep learning has shown outstanding performance in identifying intricate structures in complex high-dimensional data, especially in the domain of computer vision. The application of deep learning to early detection and automated classification of Alzheimer's disease (AD) has recently gained considerable attention, as rapid progress in neuroimaging techniques has generated large-scale multimodal neuroimaging data. A systematic review of publications using deep learning approaches and neuroimaging data for diagnostic classification of AD was performed. A PubMed and Google Scholar search was used to identify deep learning papers on AD published between January 2013 and July 2018. These papers were reviewed, evaluated, and classified by algorithm and neuroimaging type, and the findings were summarized. Of 16 studies meeting full inclusion criteria, 4 used a combination of deep learning and traditional machine learning approaches, and 12 used only deep learning approaches. The combination of traditional machine learning for classification and stacked auto-encoder (SAE) for feature selection produced accuracies of up to 98.8% for AD classification and 83.7% for prediction of conversion from mild cognitive impairment (MCI), a prodromal stage of AD, to AD. Deep learning approaches, such as convolutional neural network (CNN) or recurrent neural network (RNN), that use neuroimaging data without preprocessing for feature selection have yielded accuracies of up to 96.0% for AD classification and 84.2% for MCI conversion prediction. The best classification performance was obtained when multimodal neuroimaging and fluid biomarkers were combined. AD research that uses deep learning is still evolving, improving performance by incorporating additional hybrid data types, increasing transparency with explainable approaches that add knowledge of specific disease-related features and mechanisms.
1 Introduction
Alzheimer’s disease motivates early-detection research, while conventional neuroimaging machine learning requires multiple expert-designed preprocessing stages. This review examines deep learning architectures and their application to AD classification and progression prediction.
- Motivation: AD is a major healthcare challenge characterized by progressive cognitive decline and no validated disease-modifying treatment.The paper links this unmet need to efforts toward early detection.
- Limitations of conventional methods: Conventional neuroimaging machine learning requires feature extraction, selection, dimensionality reduction, and classifier selection.These stages require specialized knowledge, repeated optimization, and can hinder reproducibility.
- Deep learning rationale: Deep learning generates features from raw neuroimaging data through on-the-fly learning, reducing reliance on predefined feature-selection procedures.CNNs are highlighted as deep learning methods that have outperformed existing machine learning methods.
- Review scope: The review systematically searched PubMed and Google Scholar for AD deep learning studies published between January 2013 and July 2018.Studies were evaluated and classified by algorithm and neuroimaging type.
- Architectures: Deep learning methods discussed for AD include DNN, RBM, DBM, DBN, AE, sparse AE, and stacked AE.These approaches were used for AD-versus-control classification and MCI-to-AD conversion prediction with multimodal neuroimaging.
3. 1 Identification
The review screened literature systematically and retained 16 AD studies using neuroimaging and deep learning. Results compare hybrid and purely deep learning approaches across modalities, with multimodal imaging generally producing higher accuracies.
- Identification: 16 articles were included in the systematic review from 389 Google Scholar and PubMed records.The search yielded 358 Google Scholar results and 31 PubMed results.
- Identification: 25 overlapping records were removed, followed by exclusions for publication type, inaccessible full text, nonuse of deep learning, unclear neuroimaging, or missing performance results.The final 16 papers all used ADNI data in common.
- Hybrid methods: 4 studies combined deep learning for neuroimaging feature selection with traditional machine learning classifiers such as SVM.Hybrid approaches were used to tune input data for classification.
- Deep learning methods: 12 studies used only deep learning for diagnostic classification or MCI-to-AD conversion prediction.These studies included stacked auto-encoders, DNNs, and CNNs applied to neuroimaging data.
- Neuroimaging modalities: Two or more neuroimaging modalities produced higher accuracies than a single neuroimaging technique.MRI and PET were commonly combined to represent structural, amyloid, and glucose-metabolism biomarkers.
- Algorithm comparison: Limited neuroimaging data favored hybrid methods, while 3D CNNs using multimodal PET images showed better performance for both reported tasks.Deep learning approaches generally require massive amounts of data to reach desired accuracy.
5 Discussion
The review finds strong but still evolving deep-learning performance for AD classification and MCI-to-AD prediction, alongside persistent challenges in transparency, reproducibility, data integration, and generalization.
- Performance: 98.8% was the highest reported accuracy for AD classification among hybrid methods using limited neuroimaging datasets.Hybrid approaches combined traditional machine learning with deep learning components.
- Advantages and trade-offs: Deep learning automatically extracts features from input data, reducing reliance on expert-designed preprocessing and feature selection.This flexibility supports data-driven analysis but makes the extracted features difficult to identify.
- Transparency: 3D multimodal medical-image models remain difficult to interpret because nonlinear convolutional and pooling layers obscure feature importance in the original data space.This limits clarity about how anatomical information contributes to diagnostic predictions.
- Reproducibility: Performance sensitivity to random seeds and tuned hyperparameters can make deep-learning studies difficult to reproduce when configurations and code are not fully specified.The review emphasizes the need for consistent experimental settings and independent datasets.
- Future directions: Deep-learning systems have difficulty integrating heterogeneous inputs such as neuroimaging and genetic data within a single closed network.The review describes hybrid approaches that process additional information separately before combining results.
9 Funding
The review received partial support from National Institutes of Health grants and analyzed datasets funded by NIH, the Department of Defense, and private partners.
- Funding: The review was supported in part by grants from the National Institutes of Health.The listed grants include P30 AG10133, R01 AG19771, R01 AG057739, R01 CA129769, NLM R01 LM012535, and NIA R03 AG054936.
- Funding: Many reviewed studies analyzed data from the Alzheimer’s Disease Neuroimaging Initiative.The initiative was funded by the National Institutes of Health and Department of Defense.
- Funding: The Alzheimer’s Disease Neuroimaging Initiative also received support from a consortium of private partners.
11 Tables
The paper includes tables defining acronyms and summarizing 16 previously reviewed studies.
- Table 1 defines the acronyms used in the paper.
- The tables provide reference material for interpreting the review's methods and findings.
- Table 2 summarizes 16 previous studies reviewed systematically.
SEN SPE AD cMCI ncMCI NC Total
The supplied table fragment lists sensitivity and specificity notation alongside diagnostic group abbreviations and total counts.
- The table fragment includes AD, cMCI, nMCI, NC, and Total categories with accompanying numeric entries.
- SEN is defined as TP/(TP + FN).
- SPE is defined as TN/(TN + FP).
- TP, TN, FP, and FN denote true positive, true negative, false positive, and false negative, respectively.
12 Figures
The figures illustrate multilayer perceptron optimization, activation functions, deep-learning architectures, study selection, method comparisons, and neuroimaging-resource performance.
- Figure 1: Figure 1 depicts weights updated by back-propagation after least-squares error calculation until the differential value becomes 0.
- Figure 2: Figure 2 compares common deep-learning activation functions with their derivatives.
- Figure 3: Figure 3 presents architectural structures including RBM, DBM, DBN, CNN, AE, Sparse AE, and Stacked AE.
- Figure 4: Figure 4 is a PRISMA flow chart showing that 16 articles were included from 389 Google Scholar and PubMed hits.
- Figure 5: Figure 5 compares four hybrid-method studies with 12 studies using only deep learning methods.
- Figure 6: Figure 6 compares accuracy across MRI, FDG-PET, combined MRI and FDG-PET, and combined amyloid PET and FDG-PET resources.
Supplementary Material
The paper is titled Deep Learning in Alzheimer’s Disease: Diagnostic Classification and Prognostic Prediction using Neuroimaging Data and lists its authors and institutional affiliations.
- The paper concerns diagnostic classification and prognostic prediction in Alzheimer’s disease using neuroimaging data.
- Taeho Jo, Kwangsik Nho, and Andrew J. Saykin are listed as authors.
- The authors are affiliated with the Center for Neuroimaging, Department of Radiology and Imaging Sciences, Indiana University School of Medicine.
- The listed affiliations also include the Indiana Alzheimer Disease Center and Indiana University Network Science Institute.
Supplement 1. Weights calculation in the backpropagation
The backpropagation calculation starts with output error and applies the chain rule to obtain weight gradients. Weights are then updated iteratively until the differential value reaches zero.
- Weights are updated repeatedly after least-squares error calculation until the differential value becomes zero, indicating no further weight change.
- The output error is computed as the sum of errors for yo1 and yo2, while yt1 and yt2 are known target constants.
- The partial derivative with respect to w31 is decomposed into three chain-rule terms involving output error, activation, and network input.
- The output activation derivative contributes yo1∙(1-yo1) when the sigmoid function is used.
- Because Net3 includes w31yh1, its partial derivative with respect to w31 is yh1.
- Hidden-layer weights are also updated from ErrorYout because hidden activations are not directly exposed.
Supplement 2. Advanced gradient descent methods
The supplement describes momentum-based and adaptive gradient-descent methods, then lists reviewed-study configurations spanning MRI, PET, classifiers, and preprocessing strategies. The table includes multimodal models with reported accuracies up to 91.14 in the listed entries.
- Nesterov Momentum: Nesterov Momentum adds γv(t-1) to Momentum SGD to advance movement in the direction of the gradient.
- Adaptive methods: Adagrad adjusts the learning rate according to the number of variable updates.
- Adaptive methods: RMSprop adjusts the ratio between the previous value and the modified value.
- Adaptive methods: Adam combines momentum SGD and RMSprop, using Gt as the sum of squared modified gradients.
- Adaptive methods: ε is a very small constant that prevents division by zero in the optimization expression.
- Reviewed-study results: 91.14 accuracy is reported for an MRI,PET SAE with 3D CNN and softmax configuration.