Surface-based Middle-Fusion Attention model for Early Alzheimer's Disease Diagnosis
Poster No:
1437
Submission Type:
Abstract Submission
Authors:
Quan Duong1, Jin Kyu Gahm1, Dat Tran2
Institutions:
1Pusan National University, Busan, Busan, 2Pusan National University, N/A
First Author:
Co-Author(s):
Introduction:
Surface-based analysis offers many advantages for early Alzheimer's disease diagnosis. Cortical surface representaion helps analyzing cortical atrophy patterns through quantifiable metrics such as cortical thickness, sulci depth and curvature. It also reveals regional changes in cortical functions by analyzing the PET standardized uptake value ratio. Many studies deployed graph convolutional neural network to extract features from cortical surface. However, they did not achieve state-of-the-art results, possibly due to suboptimal GCNN or not incorporating multimodal features. In this study, we propose a new middle-fusion attention model that effectively leveragings multimodal cortical surface features derived from T1w MRI and FDG PET. The proposed model achieves high performances on the ADNI1, ADNI2 & ADNI3 dataset.
Methods:
1. Data Acquisition
We acquired publicly available ADNI1 datasets from the Alzheimer's disease neuroimaging initiative (ADNI). Baseline FDG PET scans were acquired for ADNI1, Amyloid PET and Tau PET for ADNI2&3 along with T1w MRI that was taken no more than two months prior. The acquired ADNI1 dataset consists of 101 CN, 208 MCI and 84 AD. ADNI2 & ADNI3 datasets comprises 258 CN, 159 MCI and 55 AD.
2. Data Preprocessing
We used FreeSurfer on T1w MRI to reconstruct cortical surface representation and generate cortical structure related metrics including cortical thickness, sulci depth and curvature. Subsequently, we employed PETSurfer on PET images as follows: registering to anatomical space, perform partial volume correction, compute SUVR and sampling SUVR onto the cortical surface. Finally, we resampled cortical thickness, sulci depth, curvature and PET SUVR onto a sixth-ordered icosphere and extract non-overlapping triangular patches.
3. Model Architecture
We designed middle-fusion attention model for effectively analyzing multimodal cortical features. The model consists of two stages: Modality-specific analysis and Inter-modality analysis. The first stage analyzes extract features from each modality through self-attention mechanism. Then features extracted from each modality are fed into the second stage, where cross-attention is performed to extract inter-modality relationship. The outputs of the second stage are concatenated and analyzed by a classifier to produce class probabilities.
We acquired publicly available ADNI1 datasets from the Alzheimer's disease neuroimaging initiative (ADNI). Baseline FDG PET scans were acquired for ADNI1, Amyloid PET and Tau PET for ADNI2&3 along with T1w MRI that was taken no more than two months prior. The acquired ADNI1 dataset consists of 101 CN, 208 MCI and 84 AD. ADNI2 & ADNI3 datasets comprises 258 CN, 159 MCI and 55 AD.
2. Data Preprocessing
We used FreeSurfer on T1w MRI to reconstruct cortical surface representation and generate cortical structure related metrics including cortical thickness, sulci depth and curvature. Subsequently, we employed PETSurfer on PET images as follows: registering to anatomical space, perform partial volume correction, compute SUVR and sampling SUVR onto the cortical surface. Finally, we resampled cortical thickness, sulci depth, curvature and PET SUVR onto a sixth-ordered icosphere and extract non-overlapping triangular patches.
3. Model Architecture
We designed middle-fusion attention model for effectively analyzing multimodal cortical features. The model consists of two stages: Modality-specific analysis and Inter-modality analysis. The first stage analyzes extract features from each modality through self-attention mechanism. Then features extracted from each modality are fed into the second stage, where cross-attention is performed to extract inter-modality relationship. The outputs of the second stage are concatenated and analyzed by a classifier to produce class probabilities.
Results:
We evaluated the proposed model using 5-fold stratified cross-validation. Area under the ROC curve was used as the evaluation metric. For ADNI1 dataset, the model achieved 97% AUC and 81% AUC on AD diagnosis (CN vs. AD) and early AD diagnosis (CN vs. MCI), respectively. For ADNI2 & ADNI3 datasets, the model scored 95% AUC on AD diagnosis and 79% AUC on early AD diagnosis.
Conclusions:
We developed a new middle-fusion attention model for early AD diagnosis. The proposed model is capable of leveraging multimodal cortical features, demonstrated by high performances on both AD diagnosis and early AD diagnosis across ADNI1, ADNI2 & ADNI3 datasets.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 2
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Image Registration and Computational Anatomy
PET Modeling and Analysis
Novel Imaging Acquisition Methods:
Multi-Modal Imaging
Keywords:
Cortex
Degenerative Disease
Machine Learning
Modeling
MRI
Positron Emission Tomography (PET)
1|2Indicates the priority used for review
Provide references using author date format
arXiv:2010.11929.
B. Fischl, "FreeSurfer," Neuroimage, vol. 62, no. 2, pp. 774-81, Aug 15 2012
D. N. Greve, "Different partial volume correction methods lead to different conclusions: an 18F-FDG-PET study of
aging," Neuroimage, vol. 132, pp. 334-343, 2016
D. N. Greve, "Cortical surface-based analysis reduces bias and variance in kinetic modeling of brain PET data,"
Neuroimage, vol. 92, pp. 225-236, 2014.
Petersen, Ronald Carl, et al. "Alzheimer's disease neuroimaging initiative (ADNI): clinical characterization." Neurology 74.3 (2010): 201-209.