Source-linked AI summary
Diagnosis of Alzheimer's Disease via Multi-modality 3D Convolutional Neural Network
Yechong Huang, Jiahang Xu, Yuncheng Zhou, Tong Tong, Xiahai Zhuang, the Alzheimer's Disease Neuroimaging Initiative
TL;DR
The paper addresses automated Alzheimer’s diagnosis and early discrimination of progressive from stable MCI using medical imaging around the hippocampal area. It proposes a 3D CNN framework that fuses T1-MRI and FDG-PET, achieving reported accuracies of 90.10% for NL/AD, 87.46% for NL/pMCI, and 76.90% for sMCI/pMCI. The authors conclude that segmentation is unnecessary for CNN classification and that a properly selected hippocampal ROI can provide diagnostic reference.
Problem
The work targets accurate, automatic discrimination of AD or potential AD from normal subjects and of sMCI from pMCI for early diagnosis.
Method
A multi-modality CNN uses paired T1-weighted MRI and 18F-FDG-PET images around the hippocampal area, with cropped ROI datasets prepared for classification.
Results
The reported accuracies were 90.10% for NL/AD, 87.46% for NL/pMCI, and 76.90% for sMCI/pMCI.
Takeaways & Limitations
The experiments indicate that segmentation is not necessary for CNN classification and that a small, properly selected hippocampal ROI can provide diagnostic reference.
Takeaways & Limitations
The study used only T1-MRI and FDG-PET, and CNN-extracted features were difficult for humans to comprehend.
Abstract
from arXiv · showhide
Alzheimer's Disease (AD) is one of the most concerned neurodegenerative diseases. In the last decade, studies on AD diagnosis attached great significance to artificial intelligence (AI)-based diagnostic algorithms. Among the diverse modality imaging data, T1-weighted MRI and 18F-FDGPET are widely researched for this task. In this paper, we propose a novel convolutional neural network (CNN) to fuse the multi-modality information including T1-MRI and FDG-PDT images around the hippocampal area for the diagnosis of AD. Different from the traditional machine learning algorithms, this method does not require manually extracted features, and utilizes the stateof-art 3D image-processing CNNs to learn features for the diagnosis and prognosis of AD. To validate the performance of the proposed network, we trained the classifier with paired T1-MRI and FDG-PET images using the ADNI datasets, including 731 Normal (NL) subjects, 647 AD subjects, 441 stable MCI (sMCI) subjects and 326 progressive MCI (pMCI) subjects. We obtained the maximal accuracies of 90.10% for NL/AD task, 87.46% for NL/pMCI task, and 76.90% for sMCI/pMCI task. The proposed framework yields comparative results against state-of-the-art approaches. Moreover, the experimental results have demonstrated that (1) segmentation is not a prerequisite by using CNN, (2) the hippocampal area provides enough information to give a reference to AD diagnosis. Keywords: Alzheimer's Disease, Multi-modality, Image Classification, CNN, Deep Learning, Hippocampal
1 School of Data Science, Fudan University, Shanghai, China
The paper concerns Alzheimer’s Disease, multi-modality imaging, CNN-based image classification, deep learning, and the hippocampal area.
- Alzheimer’s Disease is the paper’s clinical focus.
- The study uses multi-modality imaging for image classification.
- CNNs and deep learning are central methodological themes.
- The hippocampal area is a key anatomical focus.
1. Introduction
The introduction motivates automated Alzheimer’s diagnosis from hippocampal-area medical images, especially by combining MRI and PET with deep learning. It frames the work around early discrimination of AD and progressive MCI while reducing reliance on manually engineered features and segmentation.
- Alzheimer’s disease is a major neurodegenerative disease associated with dementia, cognitive impairment, and behavioral problems.
- T1-weighted MRI provides structural information, whereas 18F-FDG-PET measures regional cerebral glucose metabolism for AD assessment.
- MCI is divided retrospectively into progressive and stable forms, whose distinction supports early dementia diagnosis and therapeutic intervention.
- The work aims to distinguish AD or potential AD from normal subjects using hippocampal-area images and to classify sMCI versus pMCI for preclinical diagnosis.
- Earlier approaches commonly used preprocessing, segmentation, manual feature extraction, and classification, while deep learning can learn features automatically.
- The proposed classifier fuses MRI and PET around the hippocampal area for NL/AD, NL/pMCI, and sMCI/pMCI classification.
2. Material
The study uses ADNI MRI and PET data focused on hippocampal regions to classify AD, normal subjects, and MCI progression. It compares processing choices including segmentation, paired-modality construction, and ROI cropping.
- Region of interest: Hippocampal regions were selected as the MRI and PET region of interest because hippocampal volume is an established AD biomarker.The ROI was also motivated by reported PET metabolic changes in cortical regions relevant to AD diagnosis.
- Datasets: The Paired dataset combined MRI and FDG-PET images acquired within one year and manually verified for matching diagnoses.After filtering, it contained 647 AD, 1391 MCI, and 731 NL subjects.
- Datasets: MCI was labeled progressive when it developed to AD within three years and stable when it did not develop to AD for at least three years.Subjects lacking more than three years of follow-up were ignored; 326 pMCI and 441 sMCI subjects were identified.
- Preprocessing: MRI images were cropped around hippocampal centers into 96 × 96 × 48-voxel ROIs, with Raw, WithSeg, and Bin treatments used to assess segmentation effects.Raw retained MRI information from the hippocampus and surrounding areas, while WithSeg used binary-mask-based MRI inputs.
- Preprocessing: PET patches were generated either with MRI-matched spacing in the Origin dataset or 2 × 2 × 2 mm3 spacing in the Dilated dataset.Both PET variants used the corresponding MRI ROI center and orientation.
- Preprocessing: Data were split by patient ID into 70% training, 10% validation, and 20% testing sets.Patient-level splitting prevented subjects from the same patient appearing across multiple partitions.
3. Methodology
The methodology uses CNNs to learn image features automatically for classification, with VGG architectures providing deeper convolutional processing. VGG16 progressively reduces spatial dimensionality before fully connected and Softmax classification.
- CNN classification: CNNs automatically learn radiological image features instead of relying on manually extracted features.Their convolution kernels share weights, reducing connections and parameters relative to other deep feedforward networks.
- VGG architecture: VGG16 contains 13 convolution layers, 5 max pooling layers, and 3 fully connected layers before Softmax classification.For a 224 × 224 input, repeated convolution and pooling produce a 7 × 7 × 512 feature map.
- VGG architecture: VGG networks deepen CNN architectures by adding convolutional and pooling layers to improve large-scale image classification accuracy.The paper presents VGG as an improved AlexNet-based architecture.
4. Experiments
The experiments evaluated data representations, PET regions of interest, and single- versus multi-modality CNN designs for Alzheimer’s classification. Results favored raw MRI inputs, aligned PET regions, and multi-modality models, with protocol performance varying across tasks.
- Experimental design: The experiments tested segmentation, PET ROI spacing, and single- versus multi-modality classification protocols.The study compared datasets with and without segmentation, different PET spacings and regions, and several classifier architectures.
- PET ROI Determination: Small-spacing, small-region PET inputs achieved similar classification performance to large-spacing, large-region inputs.The PET ROI analysis considered computational limits and compared Origin and Dilated datasets with different spatial coverage and resolution.
- Influence of Segmentation: Raw MRI inputs performed best, followed by masked MRI, while binary hippocampal labels remained predictive but inferior.The results indicate CNNs can learn useful features without voxel labeling, and regions outside the hippocampi add information for separating AD from normal subjects.
- PET ROI Determination: The selected PET ROI matched the MRI ROI spatially because aligned paired images could better support locating spatial features.The Origin dataset used voxelwise alignment with MRI, whereas the Dilated dataset covered both temporal lobes at lower spatial resolution.
- Multi-modality AD Classifier: Multi-modality classifiers outperformed single-modality classifiers on the NL/AD task.Among multi-modality models, B1 had the highest accuracy and sensitivity, while B2 had the highest specificity and AUC.
- Multi-modality AD Classifier: B2 was superior on NL/pMCI and sMCI/pMCI tasks, although B1 performed slightly better on NL/AD.The models were trained with the NL/AD training set and evaluated on the NL/pMCI and sMCI/pMCI testing sets.
5. Discussion and Conclusion
The study uses a multi-modality VGG-like CNN with hippocampal T1-MRI and FDG-PET regions for AD diagnosis and prognosis. Results support segmentation-free CNN classification, transfer from NL/AD models to MCI tasks, and the use of small high-resolution hippocampal ROIs, while interpretability and modality coverage remain future-work boundaries.
- The proposed framework combines T1-MRI and FDG-PET inputs in a VGG-like CNN for AD diagnosis and prognosis.Only hippocampal areas were used as regions of interest.
- The experiments indicate that segmentation is unnecessary for CNN classification, although registration remains necessary for images from different facilities with differing spacings and orientations.The conclusion contrasts CNN processing with traditional machine-learning methods.
- Protocol B1 performed best in NL/AD classification by giving MRI and PET separate inputs while sharing convolutional kernels.The authors suggest shared weights helped extract common features across modalities.
- The NL/AD-trained model achieved 76.90% accuracy for sMCI/pMCI and 87.46% for NL/pMCI, about 5 percentage points above models trained on those tasks directly.The authors attribute this pattern to greater feature separation between NL and AD than between sMCI and pMCI.
- The study used only T1-MRI and FDG-PET, and CNN-extracted features remain difficult for humans to comprehend.Future work could add modalities and use attention mechanisms to visualize activation maps and investigate biomarkers.
- 90.10% accuracy was reached for the NL/AD task, 87.46% for NL/pMCI, and 76.90% for sMCI/pMCI.These are the reported maximal model accuracies.
- Using only hippocampal ROIs, the study indicates that small, high-resolution regions can provide valuable reference information for AD diagnosis.The authors propose deeper networks and higher-resolution medical images as future directions.