A Siamese-Transport Domain Adaptation Framework for 3D MRI Classification of Gliomas and Alzheimer’s Diseases

Neurology Computer Science Medical Laboratory Science Information Science Artificial Intelligence Medicine Oncology Medical Imaging
3D MRI Gliomas Alzheimer's Disease Domain Adaptation Deep Learning

Classification of 3D MRI Gliomas and Alzheimer’s Disease Based on the Siamese-Transport Domain Adaptation Framework

Background

In computer-aided diagnosis, 3D magnetic resonance imaging (MRI) screening plays a vital role in the early diagnosis of various brain diseases, effectively preventing the deterioration of the condition. Glioma is a common malignant brain tumor, and its treatment strategies vary depending on the tumor grade. Therefore, accurate and efficient 3D MRI classification is crucial in medical imaging analysis. However, traditional deep learning models perform poorly when applied to unlabeled data obtained clinically, mainly due to inter-domain inconsistencies, such as differences in equipment types and data acquisition parameters. Existing methods primarily focus on reducing inter-domain differences but overlook the entanglement of semantic features and domain information.

Paper Source

This paper is written by Luyue Yu, Ju Liu, Qiang Wu, Jing Wang, and Aixi Qu from Shandong University, and it is published in the January 2024 issue of IEEE Journal of Biomedical and Health Informatics. This research is funded by the Shandong Province Natural Science Foundation and the Shandong Province Key Innovation Projects.

Research Methods

Overall Workflow

The authors propose a Siamese-Transport domain adaptation framework (STDA), combining optimal transport theory and contrastive learning, for automatic 3D MRI classification and glioma multi-grade prediction. The study consists of the following main steps:

  1. Siamese-Transport Network: Designed a Siamese-based transport network that uses cross-updated parameters to extract invariant features.
  2. Optimal Cost Transport Strategy (OCTS): Proposed a strategy based on distance probability distribution to separate semantic features and domain information in the projection space.
  3. Mutual Invariance Constraint (MIC): Used mutual gradient self-training in a shared projection space to further constrain feature clustering in cross-domain tasks.

Specific Steps and Details

  1. Network Architecture Design: The STDA framework uses two identical Siamese networks that share training parameters. Depending on the classification task, the backbone network can be any deep convolutional neural network (like ResNet or VGG). In the forward pass, the two networks share training parameters; in the backward pass, parameters are updated using the gradient back-propagation algorithm based on cross-entropy and contrastive loss functions.

  2. Feature Distribution Update: Using OCTS, the optimal transport theory calculates the cost of moving from one distribution to another. It finds the invariance of potential classification features by collaboratively correcting feature mapping and updating feature distribution. In practice, the optimization of feature mapping and distribution updates is accomplished by training the feature extraction network and minimizing the loss function.

  3. Mutual Invariance Constraint: A self-supervised learning method is adopted, where mutual gradient self-training constrains the invariance of features in the projection space of the source and target domains. The stop-gradient operation is used to avoid invalid feature transfer between different domains, thereby stabilizing the feature representation during training.

Data and Experimental Design

  1. Data Sources:

    • BRATS18 Dataset: Contains multimodal MRI scans of high-grade glioma (HGG) and low-grade glioma (LGG).
    • TCGA-Brain Dataset: From The Cancer Genome Atlas project, including images of HGG and LGG samples.
    • ADNI Dataset: Contains structural MR images of Alzheimer’s disease (AD) and normal cognition (NC).
    • OASIS Dataset: Contains small samples of AD and NC data.

  2. Experimental Setup:

    • Use VGG-16 as the backbone network.
    • During training, the stochastic gradient descent (SGD) algorithm is chosen for optimization.
    • In all experiments, to ensure fair comparison, the training strategies and module calculations of all UDA methods are identical.

  3. Experimental Comparisons:

    • The experimental design compares different classical UDA methods, such as Deep Coral, JAN, and some of the latest methods.
    • Conducted ablation experiments to validate the effectiveness of each module in the Siamese-Transport network, OCTS, and MIC.

Research Results

Glioma Binary Classification Experiment

The experimental results show that the STDA method performs the best in the glioma binary classification task. Specifically, in various glioma classification tasks, STDA outperforms other UDA methods in classification accuracy and AUC values, as shown in Tables 2 and 3.

Experiments also reveal that UDA methods can improve classification accuracy even when the differences in distributions between the source and target domains are not significant. Specifically, using the optimal transport method OTDA (without the Siamese-Transport module) improved accuracy by 12.7% over models without transfer learning.

Glioma Grading Experiment

In fine-grained glioma multi-grade classification experiments, the STDA method also demonstrates the best classification performance. Compared to other UDA methods, the accuracy increased by 20.7 percentage points, indicating the strong domain adaptation capability of STDA in complex tasks.

Alzheimer’s Disease Classification Experiment

The results of the Alzheimer’s disease classification experiment (Table 4) show that the STDA method also exhibits high generalization capability across different public medical image classification datasets. Although the AUC value is slightly lower than the DAN method due to imbalanced categories within the datasets, the classification accuracy remains the highest.

Ablation Experiment Analysis

  • Siamese-Transport Network: The ablation experiments verify the effectiveness of the Siamese-Transport network architecture in feature extraction. STDA provides more stable feature representations during training.
  • Optimal Cost Transport Strategy: The OCTS method is significantly effective in correcting cross-domain distributions, especially in high-dimensional feature spaces where it can substantially reduce inter-domain distribution differences.
  • Mutual Invariance Constraint: The MIC module further stabilizes the invariant feature extraction capability of the dual network structure during training, enhancing classification accuracy.

Conclusion and Significance

The STDA framework proposed in this paper effectively addresses the small sample 3D MRI classification problem by improving feature distribution consistency and learning invariant features. It demonstrates superiority in glioma and Alzheimer’s disease classification tasks. The STDA framework is not only widely applicable but can also be further used for multi-task learning, such as brain tumor molecular markers and survival cycle prediction, meeting the needs of unbalanced small sample learning. Future research could explore more optimized data enhancement techniques, such as self-supervised learning and generative adversarial networks, to further improve the stability and accuracy of this method.

Through this study, we have not only expanded the application of 3D MRI in medical imaging analysis but also provided an effective method for addressing domain differences, offering new technical support for clinical diagnosis and treatment.