ADNI Dataset Research Articles

Anatomical Landmark Detection in 3d MRI Scan using Deep Neuro-Dynamic Programming

Automatic and precise biometric measures are crucial for diagnosis and supporting important decisions in medical imaging. For several applications of medical image analysis, automatically identifying anatomical landmarks is an important step. In this study, we determine the effectiveness of deep-neurodynamic programming methods for developing agents that can consistently and accurately recognize ground truth landmarks in medical imaging. Through interaction with an environment of three-dimensional (3D) scan ultrasound images, a simulated reinforcement learning (RL) agent learns to identify the optimum path leading to the landmark. Deep Q-Network (DQN) algorithm is applied to magnetic resonance imaging (MRI) images of the human brain. The model’s effectiveness and localization error are measured in terms of the mean Euclidean distance between the detected and ground truth landmarks’ position. We achieved a low localization error of 3.55mm during our experiments with the ADNI dataset.

DMA-HPCNet: Dual Multi-level Attention Hybrid Pyramid Convolution Neural Network for Alzheimer's Disease Classification.

Computer-aided diagnosis (CAD) plays a crucial role in the clinical application of Alzheimer's disease (AD). In particular, convolutional neural network (CNN)-based methods are highly sensitive to subtle changes caused by brain atrophy in medical images (e.g., magnetic resonance imaging, MRI). Due to computational resource constraints, most CAD methods focus on quantitative features in specific regions, neglecting the holistic nature of the images, which poses a challenge for a comprehensive understanding of pathological changes in AD. To address this issue, we propose a lightweight dual multi-level hybrid pyramid convolutional neural network (DMA-HPCNet) to aid clinical diagnosis of AD. Specifically, we introduced ResNet as the backbone network and modularly extended the hybrid pyramid convolution (HPC) block and the dual multi-level attention (DMA) module. Among them, the HPC block is designed to enhance the acquisition of information at different scales, and the DMA module is proposed to sequentially extract different local and global representations from the channel and spatial domains. Our proposed DMA-HPCNet method was evaluated on baseline MRI slices of 443 subjects from the ADNI dataset. Experimental results show that our proposed DMA-HPCNet model performs efficiently in AD-related classification tasks with low computational cost.

MAS-CL: An End-to-end Multi-atlas Supervised Contrastive Learning Framework for Brain ROI Segmentation.

Brain region-of-interest (ROI) segmentation with magnetic resonance (MR) images is a basic prerequisite step for brain analysis. The main problem with using deep learning for brain ROI segmentation is the lack of sufficient annotated data. To address this issue, in this paper, we propose a simple multi-atlas supervised contrastive learning framework (MAS-CL) for brain ROI segmentation with MR images in an end-to-end manner. Specifically, our MAS-CL framework mainly consists of two steps, including 1) a multi-atlas supervised contrastive learning method to learn the latent representation using a limited amount of voxel-level labeling brain MR images, and 2) brain ROI segmentation based on the pre-trained backbone using our MSA-CL method. Specifically, different from traditional contrastive learning, in our proposed method, we use multi-atlas supervised information to pre-train the backbone for learning the latent representation of input MR image, i.e., the correlation of each sample pair is defined by using the label maps of input MR image and atlas images. Then, we extend the pre-trained back-bone to segment brain ROI with MR images. We perform our proposed MAS-CL framework with five segmentation methods on LONI-LPBA40, IXI, OASIS, ADNI, and CC359 datasets for brain ROI segmentation with MR images. Various experimental results suggested that our proposed MAS-CL framework can significantly improve the segmentation performance on these five datasets.

Comparison of the diagnostic accuracy of resting-state fMRI driven machine learning algorithms in the detection of mild cognitive impairment

Mild cognitive impairment (MCI) is a potential therapeutic window in the prevention of dementia; however, automated detection of early cognitive deterioration is an unresolved issue. The aim of our study was to compare various classification approaches to differentiate MCI patients from healthy controls, based on rs-fMRI data, using machine learning (ML) algorithms. Own dataset (from two centers) and ADNI database were used during the analysis. Three fMRI parameters were applied in five feature selection algorithms: local correlation, intrinsic connectivity, and fractional amplitude of low frequency fluctuations. Support vector machine (SVM) and random forest (RF) methods were applied for classification. We achieved a relatively wide range of 78–87% accuracy for the various feature selection methods with SVM combining the three rs-fMRI parameters. In the ADNI datasets case we can also see even 90% accuracy scores. RF provided a more harmonized result among the feature selection algorithms in both datasets with 80–84% accuracy for our local and 74–82% for the ADNI database. Despite some lower performance metrics of some algorithms, most of the results were positive and could be seen in two unrelated datasets which increase the validity of our methods. Our results highlight the potential of ML-based fMRI applications for automated diagnostic techniques to recognize MCI patients.

Synthetic FDG‐PET hypometabolism sensitivity validation in AD

AbstractBackgroundThe availability of 18‐F fluorodeoxyglucose positron emission tomography (FDG‐PET) is not universal. We hypothesized that synthetically generated FDG‐PET images would be as sensitive to detecting the pattern of hypometabolism associated with AD as real images.MethodWe developed a deep learning‐based method to produce synthetic FDG‐PET images from 1,828 T1‐weighted MRI / real FDG‐PET image pairs from the ADNI dataset, and validated the technique on a further 284 image pairs. The technique generated synthetic FDG‐PET images which were then processed to compare Standardized Uptake Value Ratio (SUVR) with the pons as reference in 81 brain regions as defined in the Desikan‐Killiany‐Tourville and subcortical default FreeSurfer atlases.ResultWe tested the differences between synthetic and real FDG‐PET on 745 image pairs (205 controls, 365 mild cognitive impairment (MCI) and 175 AD)(Table 1). Correlations in SUVR values between synthetic and real FDG‐PET ranged between weak (r = 0.13) to strong (r = 0.63), with moderate results in key regions for AD (bilateral precuneus, r = 0.43; bilateral posterior cingulate, r = 0.37). There were significant between‐group (control vs MCI and control vs AD) differences in SUVR values for all regions between synthetic and real PET‐FDG (Figure 1) with synthetic FDG‐PET having lower values. Inter‐group effect sizes were not significantly different in the majority of brain regions (76/81)(Figure 2), with similar effect sizes in the right precuneus (synthetic: ‐0.98 vs original: ‐1.37), left (‐0.77 vs ‐1.0459) and right (‐0.80 vs ‐1.04) posterior cingulate, but different for the left precuneus (‐0.91 vs ‐1.33).ConclusionSynthetic images would increase patients’ accessibility to a meaningful modality for disease assessment while decreasing their exposure to radiation and resources in the health care system.

ALZHEIMER’S DEMENTIA EARLY DIAGNOSIS, CHARACTERIZATION, PROGNOSIS AND TREATMENT DECISION VIA A SOFTWARE‐AS‐MEDICAL DEVICE WITH AN ARTIFICIAL INTELLIGENT AGENT

AbstractBackgroundTRACE4AD (DeepTrace Technologies s.r.l, Italy) is a machine learning‐based software‐as‐medical device able to predict the conversion to Alzheimer’s disease (AD) dementia of subjects at risk within 24‐months exploiting automatic processing of T1‐weighted MPRAGE brain MRI study and neuropsychological tests. TRACE4AD provides a report with the predicted individual risk of conversion to AD dementia, specific cognitive deficits, and suggestions supporting neurologists in diagnosis and characterization, prognosis, and decision‐making. We tested TRACE4AD in the clinical setting in its ability, at baseline, to: a) predict amnestic Mild Cognitive Impairment (MCI) conversion to AD dementia within 24‐months; b) characterize cognitive deficits; c) support neurologists' decision‐making.MethodWe retrospectively included 92 subjects (mean age 73.12±7.6, 46% female): 32 patients from two Italian centers where TRACE4AD was implemented after user training; and 60 subjects from the ADNI dataset. All patients had a brain MPRAGE study at baseline, 77/92 patients (83.7%) also performed a neuropsychological assessment at baseline, 75/92 patients (81.5%) had a stable clinical diagnosis at 24‐month. TRACE4AD extracted the gray matter map from MPRAGE and used it (combined with cognitive measures when available) to make inferences. Reference standards were: a) the neurologist’s clinical diagnosis at 24‐months, b) the neuropsychological assessment at the baseline, c) the agreement with the neuro exam and intervention decision time and type defined by neurologists at the baseline.ResultTRACE4AD accurately predicted conversion to AD dementia in 93.3% of patients based on the MRI, and in 96.6% based on MRI and cognitive measures. Cognitive deficits were found in agreement with the neuropsychological assessment for all patients except one who presented with major depression. We found disagreement between the neurologist’s decision at baseline and the tool in only two patients, defined with normal cognition by the neurologist and predicted at high risk of AD dementia conversion by the tool. TRACE4AD supported neurologists' decision by 12 months in 15/17 patients for the prompt decision at baseline.ConclusionTRACE4AD is promising, safe, and effective in supporting neurologists in the clinical practice of MCI across different centers.

Development and comprehensive evaluation of high‐resolution 0.5 mm isotropic T2‐weighted, PD‐weighted, T2 map and FLAIR templates of the MIITRA atlas.

AbstractBackgroundVoxel‐wise, disease‐monitoring and other MRI investigations involving T2‐weighted(T2w), PD‐weighted(PDw), T2 map and FLAIR templates typically utilize young adult templates such as those of the ICBM atlas. Moreover, a comprehensive assessment of how such available standardized templates perform on aging population has not yet been performed. Here, a high‐resolution older‐adult population based T2w, PDw T2 map and FLAIR templates of the MIITRA atlas was developed using theories of super‐resolution and sparse‐representation and compared to other available standardized templates in terms of image quality, spatial normalization accuracy and representativeness of the older adult brain.MethodT1‐weighted MPRAGE of isotropic 1mm resolution, five spin‐echo(0.5×0.5×0.5 mm3,TE = 20,40,60,80,100ms) and FLAIR data (0.5×0.5×4 mm3) were collected on 3T MRI scanners for 400 non‐demented older adults (50% male; 65‐95 years of age, 54% white and 43% black) participating in the construction of the MIITRA atlas. The subject T2 maps and PD‐weighted data were computed by fitting the T2 decay equation. The template construction technique involved upsampling the images to 0.5×0.5×0.5 mm3 using non‐local upsampling, and converting the T2‐weighted(TE = 60ms) to synthetic T1‐weighted, followed by linear registration to upsampled T1‐weighted data using ANTs. Next, the resulting transformations concatenated with the combined iterative T1w and DTI driven MIITRA atlas deformations were used to forward map each participant to exact physical locations in the MIITRA space. Finally, sparse representation data fusion technique was utilized on the transformed signals to construct the templates at isotropic 0.5mm resolution (Figure.1). MIITRA third echo T2‐weighted and FLAIR template was compared to available standardized templates (Figure. 2) in terms of inter‐subject spatial normalization accuracy when used for spatial normalization of 400 ADNI T2‐weighted and FLAIR participants respectively.ResultThe T2‐weighted and FLAIR templates of the MIITRA atlas exhibited higher image sharpness exemplified by higher spatial frequency content of the normalized power spectra(Figure. 2). MIITRA templates demonstrated higher inter‐subject spatial normalization accuracy and required the least deformations for spatial normalization of ADNI data(Figure. 3)ConclusionThe new T2w, PDw, and FLAIR templates of the MIITRA atlas exhibited superior image quality, required lower deformations, facilitating higher spatial normalization accuracy for use in studies on older adults such as the ADNI datasets.

Hippocampal‐to‐ventricle ratio (HVR) is better related to age and cognition than Hippocampal Volume

AbstractBackgroundHippocampal volume (HCvol) is an important biomarker in the study of neurodegeneration (1‐3). Given its high variability (4) HCvol is commonly normalized using the intracranial volume (HCvol/ICV) (5). The Hippocampal‐to‐ventricle ratio (HVR) considers instead the nearby expanding ventricular space and has shown stronger negative association with age and positive association with delayed memory compared to HCvol in healthy aging (6).MethodUsing a Convolutional Neural Network (7) we extracted the volumes from the hippocampus and the surrounding temporal horn of the lateral ventricles to obtain the HCvol, HCvol/ICV and HVR on the baseline T1w MRI scans from the ADNI dataset (CN = 502, MCI = 814, AD = 323). We converted these measurements to Z‐scores using the mean and standard deviation of the CN cohort to facilitate comparisons. We used Pearson’s correlation coefficient to compare the relationships between each of the three Hippocampal measures with age, ADAS13, and the Immediate, Learning, Forgetting and Forgetting‐percentage metrics of the Rey’s Auditory Verbal Learning Test (RAVLT). Finally, we compared the correlations of HCvol/ICV and HVR with a bias‐corrected‐and‐accelerated bootstrap (BCa; 10,000 resamples) (8).ResultThe correlations of HCvol and HCvol/ICV were similar for age (‐.41 [‐.32, ‐.48] vs ‐.38 [‐.3, ‐.45]), ADAS13 (‐.28 [‐.19, ‐.45] vs ‐.26 [‐.17, ‐.45]) and RAVLT subtests (absolute values: .2 [0, .4] vs .2 [.01, .41]). HVR showed significantly stronger negative correlations with age compared to HCvol/ICV (Fig. 1; CN: BCa = .144 [.102‐.187]; MCI: BCa = .128 [.096‐.160]; AD: BCa = .068 [.012‐.122]). While the negative correlations between HVR and ADAS13 scores were higher than HCvol, the comparison was not statistically significant (Fig. 2). Figure 3 shows the relationships between HC measures and RAVLT subtests: HVR had stronger correlations than HCvol for CN and AD on the Immediate subtest, while HCvol’s relationship with Forgetting score was higher for MCI. None of these were statistically significant.ConclusionConsidering ventricular expansion and hippocampal shrinkage in the HVR better explains the variance related to age and cognition than using hippocampal volume by itself in healthy controls and patients with mild cognitive impairment and dementia.

Synthetic FDG‐PET hypometabolism sensitivity validation in AD

AbstractBackgroundThe availability of 18‐F fluorodeoxyglucose positron emission tomography (FDG‐PET) is not universal. We hypothesized that synthetically generated FDG‐PET images would be as sensitive to detecting the pattern of hypometabolism associated with AD as real images.MethodWe developed a deep learning‐based method to produce synthetic FDG‐PET images from 1,828 T1‐weighted MRI / real FDG‐PET image pairs from the ADNI dataset, and validated the technique on a further 284 image pairs. The technique generated synthetic FDG‐PET images which were then processed to compare Standardized Uptake Value Ratio (SUVR) with the pons as reference in 81 brain regions as defined in the Desikan‐Killiany‐Tourville and subcortical default FreeSurfer atlases.ResultWe tested the differences between synthetic and real FDG‐PET on 745 image pairs (205 controls, 365 mild cognitive impairment (MCI) and 175 AD)(Table 1). Correlations in SUVR values between synthetic and real FDG‐PET ranged between weak (r = 0.13) to strong (r = 0.63), with moderate results in key regions for AD (bilateral precuneus, r = 0.43; bilateral posterior cingulate, r = 0.37). There were significant between‐group (control vs MCI and control vs AD) differences in SUVR values for all regions between synthetic and real PET‐FDG (Figure 1) with synthetic FDG‐PET having lower values. Inter‐group effect sizes were not significantly different in the majority of brain regions (76/81)(Figure 2), with similar effect sizes in the right precuneus (synthetic: ‐0.98 vs original: ‐1.37), left (‐0.77 vs ‐1.0459) and right (‐0.80 vs ‐1.04) posterior cingulate, but different for the left precuneus (‐0.91 vs ‐1.33).ConclusionSynthetic images would increase patients’ accessibility to a meaningful modality for disease assessment while decreasing their exposure to radiation and resources in the health care system.

Disentangle First, Then Distill: A Unified Framework for Missing Modality Imputation and Alzheimer's Disease Diagnosis.

Multi-modality medical data provide complementary information, and hence have been widely explored for computer-aided AD diagnosis. However, the research is hindered by the unavoidable missing-data problem, i.e., one data modality was not acquired on some subjects due to various reasons. Although the missing data can be imputed using generative models, the imputation process may introduce unrealistic information to the classification process, leading to poor performance. In this paper, we propose the Disentangle First, Then Distill (DFTD) framework for AD diagnosis using incomplete multi-modality medical images. First, we design a region-aware disentanglement module to disentangle each image into inter-modality relevant representation and intra-modality specific representation with emphasis on disease-related regions. To progressively integrate multi-modality knowledge, we then construct an imputation-induced distillation module, in which a lateral inter-modality transition unit is created to impute representation of the missing modality. The proposed DFTD framework has been evaluated against six existing methods on an ADNI dataset with 1248 subjects. The results show that our method has superior performance in both AD-CN classification and MCI-to-AD prediction tasks, substantially over-performing all competing methods.

Federated BrainAge estimation from MRI: a proof of concept

AbstractBackgroundThe gap between chronological and biological brain‐age, represented by MRI features, associates with dementia risk (Wang et. al., He et. al.). Age‐scores are predicted, however, using deep learning models, which require large amounts of MRI data. This poses a challenge, as patient brain MRI data is highly privacy‐sensitive and legal barriers often impede the sharing or transferring of data. Federated learning provides a potential solution by enabling training a global model on distributed data, allowing biobanks to keep data local while creating a shared model. In this study, we assess whether a federated set‐up is feasible for brain‐age estimation.MethodWe set up a federated learning framework across three biobanks (Fig. 2) to train a BrainAge model (Wang et. al). Data of healthy individuals from the ADNI dataset was randomly split across the three biobanks (N = 221, 219, 217). MRI images were pre‐processed into gray matter density maps (Good et. al.). The model was iteratively trained using federated averaging (McMahan et. al.) on data at two biobanks (70/30 training/validation split), and tested at the remaining biobank. We trained the model for ten rounds, and optimized for the number of epochs, learning rate, and learning decay hyperparameters. These results were compared to a centrally trained model, where the training samples were pooled together. Additionally, we applied the trained model to individuals with Mild Cognitive Impairment and Alzheimer’s Disease and performed a survival analysis. We assessed performance using Mean Absolute Error (MAE).ResultFederated learning (MAE = 4.02) achieves similar test performance as central learning (MAE = 3.94) (Table 1). The model yielded an MAE of 4.89 for MCI and 5.45 for AD patients; this increase may be explained by their neurodegeneration. We also see a statistically different survival curve for developing AD in individuals with a brain‐age‐gap (Fig. 1), displaying an ‘older’ profile for their chronological age. This might indicate higher risk of developing AD, in line with previous results (Wang et. al.).ConclusionResults show federated learning has the potential to address the challenge of data accessibility for machine learning in complex and data hungry research, including dementia.

De-accumulated error collaborative learning framework for predicting Alzheimer’s disease progression

Alzheimer’s disease (AD) is a chronic neurodegenerative disorder where precise prediction of progression is crucial for improving clinical diagnosis. However, missing data often complicates AD prediction in clinical practice. Existing studies primarily have employed recurrent neural networks to impute missing data, but this approach suffers from the “error accumulation” problem, leading to unsatisfactory prediction results. To address this issue, this study proposes the LSTM-TSGAIN collaborative learning framework, which consists of the following three main aspects: using a generative adversarial imputation method to reduce LSTM prediction errors; collaborative training of the adversarial imputation fusion module and the time series learning module to improve the performance of the model; the model inputs were adjusted to variable lengths to accommodate the differences in the number of visits of different subjects. The effectiveness of the model is demonstrated through experiments using longitudinal data from the ADNI dataset of 1256 subjects, which shows its superiority over state-of-the-art methods.

A slice selection guided deep integrated pipeline for Alzheimer’s prediction from Structural Brain MRI

Alzheimer’s disease, a progressive form of dementia, has risen to become the fifth leading cause of death among individuals aged 65 and older. The diagnosis of Alzheimer’s is both time-consuming and costly, involving radiologists and clinical experts at multiple stages, which presents a significant challenge in the medical field. Moreover, cases of Alzheimer’s and dementia often go undiagnosed or misdiagnosed worldwide. To address this issue, medical experts meticulously analyze patients’ structural MRI (sMRI) scans to identify potential abnormalities linked to Alzheimer’s or other forms of dementia. Recognizing the devastating impact of this disease on people’s lives, Artificial Intelligence (AI) researchers have been dedicated to developing automated solutions for early-stage Alzheimer’s diagnosis in recent years, aiming to support medical practitioners in their efforts. Despite the application of various AI-driven solutions that use sMRI data for Alzheimer’s diagnosis, there are still research gaps that need attention. These gaps include the need for guided slice selection and the development of a simpler yet effective integrated pipeline where each stage of the process is fully automated, eliminating the need for medical practitioner intervention. In this study, we propose an integrated automated solution that incorporates a guided machine learning-based selection process using K-Means++ leading to a Gradient Boosting-based method for identifying the 16 most relevant 2-dimensional sMRI slices from 3-dimensional sMRI data. This step is crucial for accurate Alzheimer’s classification. Furthermore, we introduce a deep learning architecture that combines EfficientNetV2S-based transfer learning with densely-learned features in an optimized manner. To evaluate the effectiveness of our proposed deep-integrated architecture, we used two benchmark datasets from ADNI and OASIS, conducting rigorous experimental analysis and validation. The results demonstrated that our integrated architecture outperformed all other experimented architectures, achieving a 20-Fold Cross Validation Accuracy of 83.64% (CN vs AD), 82.69% (CN vs MCIc), and 71.40% (CN vs MCInc) on the ADNI dataset, and 91.54% (CN vs AD) on the OASIS dataset. This signifies the potential of our approach in improving Alzheimer’s diagnosis accuracy and offers hope for early detection and intervention in this debilitating disease.

Joint learning framework of cross-modal synthesis and diagnosis for Alzheimer’s disease by mining underlying shared modality information

Alzheimer’s disease (AD) is one of the most common neurodegenerative disorders presenting irreversible progression of cognitive impairment. How to identify AD as early as possible is critical for intervention with potential preventive measures. Among various neuroimaging modalities used to diagnose AD, functional positron emission tomography (PET) has higher sensitivity than structural magnetic resonance imaging (MRI), but it is also costlier and often not available in many hospitals. How to leverage massive unpaired unlabeled PET to improve the diagnosis performance of AD from MRI becomes rather important. To address this challenge, this paper proposes a novel joint learning framework of unsupervised cross-modal synthesis and AD diagnosis by mining underlying shared modality information, improving the AD diagnosis from MRI while synthesizing more discriminative PET images. We mine underlying shared modality information in two aspects: diversifying modality information through the cross-modal synthesis network and locating critical diagnosis-related patterns through the AD diagnosis network. First, to diversify the modality information, we propose a novel unsupervised cross-modal synthesis network, which implements the inter-conversion between 3D PET and MRI in a single model modulated by the AdaIN module. Second, to locate shared critical diagnosis-related patterns, we propose an interpretable diagnosis network based on fully 2D convolutions, which takes either 3D synthesized PET or original MRI as input. Extensive experimental results on the ADNI dataset show that our framework can synthesize more realistic images, outperform the state-of-the-art AD diagnosis methods, and have better generalization on external AIBL and NACC datasets.

Fine Perceptive GANs for Brain MR Image Super-Resolution in Wavelet Domain.

Magnetic resonance (MR) imaging plays an important role in clinical and brain exploration. However, limited by factors such as imaging hardware, scanning time, and cost, it is challenging to acquire high-resolution MR images clinically. In this article, fine perceptive generative adversarial networks (FP-GANs) are proposed to produce super-resolution (SR) MR images from the low-resolution counterparts. By adopting the divide-and-conquer scheme, FP-GANs are designed to deal with the low-frequency (LF) and high-frequency (HF) components of MR images separately and parallelly. Specifically, FP-GANs first decompose an MR image into LF global approximation and HF anatomical texture subbands in the wavelet domain. Then, each subband generative adversarial network (GAN) simultaneously concentrates on super-resolving the corresponding subband image. In generator, multiple residual-in-residual dense blocks are introduced for better feature extraction. In addition, the texture-enhancing module is designed to trade off the weight between global topology and detailed textures. Finally, the reconstruction of the whole image is considered by integrating inverse discrete wavelet transformation in FP-GANs. Comprehensive experiments on the MultiRes_7T and ADNI datasets demonstrate that the proposed model achieves finer structure recovery and outperforms the competing methods quantitatively and qualitatively. Moreover, FP-GANs further show the value by applying the SR results in classification tasks.

Diagnosis of Alzheimer’s Disease Based on Accelerated Mirror Descent Optimization and a Three-Dimensional Aggregated Residual Network

Alzheimer’s disease (AD), a neuropsychiatric disorder, continually arises in the elderly. To date, no targeted medications have been developed for AD. Early and fast diagnosis of AD plays a pivotal role in identifying potential AD patients, enabling timely medical interventions, and mitigating disease progression. Computer-aided diagnosis (CAD) becomes possible with the burgeoning of deep learning. However, the existing CAD models for processing 3D Alzheimer’s disease images usually have the problems of slow convergence, disappearance of gradient, and falling into local optimum. This makes the training of 3D diagnosis models need a lot of time, and the accuracy is often poor. In this paper, a novel 3D aggregated residual network with accelerated mirror descent optimization is proposed for diagnosing AD. First, a novel unbiased subgradient accelerated mirror descent (SAMD) optimization algorithm is proposed to speed up diagnosis network training. By optimizing the nonlinear projection process, our proposed algorithm can avoid the occurrence of the local optimum in the non-Euclidean distance metric. The most notable aspect is that, to the best of our knowledge, this is the pioneering attempt to optimize the AD diagnosis training process by improving the optimization algorithm. Then, we provide a rigorous proof of the SAMD’s convergence, and the convergence of SAMD is better than any existing gradient descent algorithms. Finally, we use our proposed SAMD algorithm to train our proposed 3D aggregated residual network architecture (ARCNN). We employed the ADNI dataset to train ARCNN diagnostic models separately for the AD vs. NC task and the sMCI vs. pMCI task, followed by testing to evaluate the disease diagnostic outcomes. The results reveal that the accuracy can be improved in diagnosing AD, and the training speed can be accelerated. Our proposed method achieves 95.4% accuracy in AD diagnosis and 79.9% accuracy in MCI diagnosis; the best results contrasted with several state-of-the-art diagnosis methods. In addition, our proposed SAMD algorithm can save about 19% of the convergence time on average in the AD diagnosis model compared with the gradient descent algorithms, which is very momentous in clinic.

APOEɛ4 Status and Plasma p-tau181 Levels May Influence Memory and Executive Function Decline in Older Adults Without Dementia.

Elevated tau phosphorylation has been linked to the Apolipoprotein E (APOE) ɛ4 allele, which is considered one of the most significant genes related to Alzheimer's disease (AD). However, it is uncertain whether the impact of increased plasma tau phosphorylated at threonine 181 (p-tau181) on memory and executive function decline would be greater among APOEɛ4 carriers. To investigate the effects of plasma p-tau181 and APOEɛ4 on memory and executive function. The longitudinal analysis included 608 older adults without dementia (aged 72±7 years; 47% female; follow-up period of 1.59±1.47 years) from the ADNI dataset, including 180 individuals with normal cognition and 429 individuals with mild cognitive impairment. Linear mixed-effects models were utilized to assess the contributions of APOEɛ4 status and plasma p-tau181 to longitudinal changes in memory composite score and executive function composite score. At baseline, the APOEɛ4+/Tau+ group exhibited poorer performance in memory composite score and executive function composite score, and an elevated load of cerebrospinal fluid Aβ and tau pathologies. To further understand longitudinal changes, we compared groups directly based on plasma p-tau181 and APOEɛ4 status (four groups: APOEɛ4-/Tau-, APOEɛ4-/Tau+, APOEɛ4+/Tau-, APOEɛ4+/Tau+). Both the memory composite score and executive function composite score showed a significantly greater decline in the APOEɛ4+/Tau+ group than in all other groups. Our findings indicate that there is an interaction between plasma p-tau181 levels and APOEɛ4 status, which contributes to the longitudinal changes of memory and executive function in older adults without dementia.

Identifying underlying patterns in Alzheimer's disease trajectory: a deep learning approach and Mendelian randomization analysis

Alzheimer's disease (AD) is a heterogeneously progressive neurodegeneration disorder with varied rates of deterioration, either between subjects or within different stages of a certain subject. Estimating the course of AD at early stages has treatment implications. We aimed to analyze disease progression to identify distinct patterns in AD trajectory. We proposed a deep learning model to identify underlying patterns in the trajectory from cognitively normal (CN) to a state of mild cognitive impairment (MCI) to AD dementia, by jointly predicting time-to-conversion and clustering out distinct subgroups characterized by comprehensive features as well as varied progression rates. We designed and validated our model on the ADNI dataset (1370 participants). Prediction of time-to-conversion in AD trajectory was used to validate the expression of the identified patterns. Causality between patterns and time-to-conversion was further inferred using Mendelian randomization (MR) analysis. External validation was performed on the AIBL dataset (233 participants). The proposed model clustered out patterns characterized by significantly different biomarkers and varied progression rates. The discovered patterns also showed a strong prediction ability, as indicated by hazard ratio (CN→MCI, HR=3.51, p<0.001; MCI→AD, HR=8.11, p<0.001), C-Index (CN→MCI, 0.618; MCI→AD, 0.718), and AUC (CN→MCI, 3 years 0.802, 5 years 0.876; MCI→AD, 3 years 0.914, 5 years 0.957). In the external validation cohort, our model demonstrated competitive performance on conversion time prediction (CN→MCI, C-Index=0.693; MCI→AD, C-Index=0.752). Moreover, suggestive associations between CN→MCI/MCI→AD patterns with four/three SNPs were mediated and MR analysis indicated a causal link between MCI→AD patterns and time-to-conversion in the first three years. Our proposed model identifies biologically and clinically meaningful patterns from real-world data and provides promising performance on time-to-conversion prediction in AD trajectory, which could promote the understanding of disease progression, facilitate clinical trial design, and provide potential for decision-making. The National Key Research and Development Program of China, the Key R&D Program of Zhejiang, and the National Nature Science Foundation of China.

PINet: Privileged Information Improve the Interpretablity and generalization of structural MRI in Alzheimer's Disease.

The irreversible and progressive atrophy by Alzheimer's Disease resulted in continuous decline in thinking and behavioral skills. To date, CNN classifiers were widely applied to assist the early diagnosis of AD and its associated abnormal structures. However, most existing black-box CNN classifiers relied heavily on the limited MRI scans, and used little domain knowledge from the previous clinical findings. In this study, we proposed a framework, named as PINet, to consider the previous domain knowledge as a Privileged Information (PI), and open the black-box in the prediction process. The input domain knowledge guides the neural network to learn representative features and introduced intepretability for further analysis. PINet used a Transformer-like fusion module Privileged Information Fusion (PIF) to iteratively calculate the correlation of the features between image features and PI features, and project the features into a latent space for classification. The Pyramid Feature Visualization (PFV) module served as a verification to highlight the significant features on the input images. PINet was suitable for neuro-imaging tasks and we demonstrated its application in Alzheimer's Disease using structural MRI scans from ADNI dataset. During the experiments, we employed the abnormal brain structures such as the Hippocampus as the PI, trained the model with the data from 1.5T scanners and tested from 3T scanners. The F1-score showed that PINet was more robust in transferring to a new dataset, with approximatedly 2% drop (from 0.9471 to 0.9231), while the baseline CNN methods had a 29% drop (from 0.8679 to 0.6154). The performance of PINet was relied on the selection of the domain knowledge as the PI. Our best model was trained under the guidance of 12 selected ROIs, major in the structures of Temporal Lobe and Occipital Lobe. In summary, PINet considered the domain knowledge as the PI to train the CNN model, and the selected PI introduced both interpretability and generalization ability to the black box CNN classifiers.

Development and validation of a deep learning-based automatic segmentation model for assessing intracranial volume: comparison with NeuroQuant, FreeSurfer, and SynthSeg.

To develop and validate a deep learning-based automatic segmentation model for assessing intracranial volume (ICV) and to compare the accuracy determined by NeuroQuant (NQ), FreeSurfer (FS), and SynthSeg. This retrospective study included 60 subjects [30 Alzheimer's disease (AD), 21 mild cognitive impairment (MCI), 9 cognitively normal (CN)] from a single tertiary hospital for the training and validation group (50:10). The test group included 40 subjects (20 AD, 10 MCI, 10 CN) from the ADNI dataset. We propose a robust ICV segmentation model based on the foundational 2D UNet architecture trained with four types of input images (both single and multimodality using scaled or unscaled T1-weighted and T2-FLAIR MR images). To compare with our model, NQ, FS, and SynthSeg were also utilized in the test group. We evaluated the model performance by measuring the Dice similarity coefficient (DSC) and average volume difference. The single-modality model trained with scaled T1-weighted images showed excellent performance with a DSC of 0.989 ± 0.002 and an average volume difference of 0.46% ± 0.38%. Our multimodality model trained with both unscaled T1-weighted and T2-FLAIR images showed similar performance with a DSC of 0.988 ± 0.002 and an average volume difference of 0.47% ± 0.35%. The overall average volume difference with our model showed relatively higher accuracy than NQ (2.15% ± 1.72%), FS (3.69% ± 2.93%), and SynthSeg (1.88% ± 1.18%). Furthermore, our model outperformed the three others in each subgroup of patients with AD, MCI, and CN subjects. Our deep learning-based automatic ICV segmentation model showed excellent performance for the automatic evaluation of ICV.