Source-linked AI summary
Multimodal and Multiscale Deep Neural Networks for the Early Diagnosis of Alzheimer's Disease using structural MR and FDG-PET images
Donghuan Lu, Karteek Popuri, Weiguang Ding, Rakesh Balachandar, Mirza Faisal Beg
TL;DR
Prodromal Alzheimer’s diagnosis is difficult because clinical diagnosis has limited specificity, while available neuroimaging modalities capture complementary brain structure and metabolism. The paper combines multiscale MRI and FDG-PET patch features in a multimodal deep neural network, achieving reported accuracy of 85.68% for subjects converting within three years and outperforming prior methods in the evaluated discrimination tasks.
Problem
Prodromal Alzheimer’s diagnosis requires more reliable detection because clinical diagnosis has limited specificity for identifying underlying pathology.
Method
The method combines patch-wise MRI and FDG-PET features across multiple scales, using six modality-scale DNNs followed by a fusion DNN.
Results
85.68% accuracy was achieved for identifying subjects who would convert to Alzheimer’s disease within three years.
Takeaways & Limitations
The approach outperformed previous methods in classifying progressive and stable MCI individuals across single-modality and multimodal imaging settings.
Takeaways & Limitations
Classification accuracy remains constrained by limited characterization of Alzheimer’s features in FDG-PET and structural MRI images.
Abstract
from arXiv · showhide
Alzheimer's Disease (AD) is a progressive neurodegenerative disease. Amnestic mild cognitive impairment (MCI) is a common first symptom before the conversion to clinical impairment where the individual becomes unable to perform activities of daily living independently. Although there is currently no treatment available, the earlier a conclusive diagnosis is made, the earlier the potential for interventions to delay or perhaps even prevent progression to full-blown AD. Neuroimaging scans acquired from MRI and metabolism images obtained by FDG-PET provide in-vivo view into the structure and function (glucose metabolism) of the living brain. It is hypothesized that combining different image modalities could better characterize the change of human brain and result in a more accuracy early diagnosis of AD. In this paper, we proposed a novel framework to discriminate normal control(NC) subjects from subjects with AD pathology (AD and NC, MCI subjects convert to AD in future). Our novel approach utilizing a multimodal and multiscale deep neural network was found to deliver a 85.68\% accuracy in the prediction of subjects within 3 years to conversion. Cross validation experiments proved that it has better discrimination ability compared with results in existing published literature.
Introduction
Reliable prodromal Alzheimer’s diagnosis remains necessary because clinical diagnosis has limited specificity, while MRI and FDG-PET provide complementary structural and functional information. The paper proposes combining multiscale, multimodal features with deep neural networks and reports improved discrimination, including 85.68% accuracy for subjects converting within three years.
- Clinical diagnosis has limited specificity for identifying prodromal Alzheimer’s disease, motivating tools that can reliably detect pathology early.
- MRI provides structural brain information, whereas FDG-PET measures resting-state glucose metabolism reflecting underlying tissue activity.
- Deep-learning neuroimaging methods face limited sample sizes relative to feature dimensionality, while patch segmentation can lose discriminative information through downsampling.
- 85.68% accuracy was achieved for identifying subjects expected to convert to Alzheimer’s disease within three years.
- The proposed approach combined multiscale and multimodal processing with a deep neural network and outperformed previous methods on the discriminative task.
Methods
The framework preprocesses paired MRI and FDG-PET scans into patch-wise structural and metabolic features, then classifies subjects with a multimodal multiscale neural network. Six scale-modality networks produce latent representations that a fusion network combines, with pre-training, regularization, and ensemble strategies used during training.
- Image preprocessing and classification: The framework first segments MRI and FDG-PET images into patches, extracts patch features, and trains a deep network to classify Alzheimer’s pathology.
- Data: The study used ADNI subjects with both T1-weighted MRI and FDG-PET scans, categorized into five clinical groups.
- Image preprocessing: MRI processing subdivided gray matter into 87 anatomical regions of interest after segmentation and quality control.
- Multiscale Deep Neural Network: Each modality and scale supplied patch-wise biomarkers to an independent DNN, whose latent features were fused by another DNN for final classification.
- Training: Each DNN underwent unsupervised stacked-autoencoder pre-training before supervised fine-tuning, followed by joint tuning of all MMDNN parameters.
- Regularization and optimization: Dropout, early stopping, and ensemble classifiers were used to reduce overfitting and improve classification stability with limited validation data.
Results and Discussion
The proposed network was evaluated against prior methods, across multiscale features and training-set compositions, for classifying AD-related progression. Multiscale fusion and adding progressive subjects improved discrimination, while imperfect image characterization limited classification accuracy.
- Compare with State-of-the-Art Methods: The proposed network outperformed state-of-the-art methods in classifying progressive and stable MCI individuals using single or multimodal imaging.
- Compare with State-of-the-Art Methods: A 3.5% accuracy advantage was reported over a domain-transfer method, despite using no auxiliary knowledge.
- Early Diagnosis: Classifier accuracy and sensitivity progressively improved when pMCI and pNC subjects were added to training with sNC and sAD.
- Multiscale Classification: Fusing multiscale features yielded superior accuracy to uniscale features, although accuracy showed no consistent trend across individual patch sizes.
- Early Diagnosis: 90.08%, 85.61%, and 81.20% accuracy were reported for identifying AD-risk individuals approximately 1, 2, and 3 years before disease onset.
- Early Diagnosis: The feature representations differed across training sets containing sNC, sAD, pMCI, and pNC samples.
- Early Diagnosis: Classification accuracy was constrained because FDG-PET and structural MRI features can overlap with conditions other than AD, including normal controls.
Conclusion
The study developed a deep neural network using multiscale, multimodal imaging patterns to identify individuals at risk for Alzheimer’s disease. The classifier distinguished Alzheimer’s pathology from stable normal controls and predicted conversion before illness onset.
- 82.93% accuracy was achieved in distinguishing individuals with Alzheimer’s pathology from stable normal controls using cross-validation.
- The classifier combined patterns from different resolutions and imaging modalities to distinguish Alzheimer’s trajectories from stable normal controls.
- Combined FDG-PET and structural MRI produced better performance than either modality alone.
- The classifier trained on pNC, pMCI, and sAD achieved the highest classification accuracy among the evaluated training samples.
- 85.68% sensitivity was obtained for recognizing individuals with Alzheimer’s pathology three years before illness onset.
Consortia
The paper acknowledges the Alzheimer’s Disease Neuroimaging Initiative contributors whose names are listed across the consortium statements.
- The consortium statement lists Michael Weiner, Paul Aisen, Ronald Petersen, and other contributors.
- The passages provide extended author and investigator listings for the Alzheimer’s Disease Neuroimaging Initiative.
- The contributor list continues with Stephen Pasternack, Irina Rachisky, Dick Trost, and additional investigators.
Author contributions statement
The authors specify individual responsibilities for network development, neuroimage processing, experimental design, result interpretation, and manuscript review.
- Donghuan Lu and Weiguang Ding built the deep neural network.
- Donghuan Lu and Karteek Popuri processed the neuroimage data.
- Rakesh Balachander and Mirza Faisal Beg designed the experiments and interpreted the results, while all authors reviewed the manuscript.
Additional information
The authors report that they have no competing interest.
- The authors declare that they have no competing interest.
- The paper includes an explicit competing-interest disclosure.
- The disclosed competing-interest status is none.