Multi-template Meta-information Regularized Network for Alzheimer’s Disease Diagnosis: A Study Based on Structural MRI

Research Background

Alzheimer’s Disease (AD) is a progressive neurodegenerative disorder, and its diagnosis and early detection is a significant challenge in the medical field. Structural MRI (sMRI) is widely used in computer-aided diagnosis of Alzheimer’s due to its capability of providing detailed brain morphological patterns and anatomical features. Although previous studies have validated the effectiveness of combining metadata (such as age, gender, and education years) for AD diagnosis using sMRI, existing methods mainly focus on the correlation or confounding effects of metadata with AD, such as gender biases and normal aging issues, making it difficult to fully exploit the impact of metadata on AD diagnosis. To address these issues, this study constructs a novel Multi-template Meta-information Regularized Network (MMRN) for AD diagnosis using sMRI.

Research Source

This paper was authored by Kanfu Han, Gang Li, Zhiwen Fang, and Yang Feng from the School of Biomedical Engineering, Southern Medical University, Guangdong Provincial Key Laboratory of Medical Image Processing, Guangdong Provincial Engineering Laboratory for Medical Imaging and Diagnostic Technology, and the University of North Carolina at Chapel Hill. The study was published on December 18, 2023 in the IEEE Transactions on Medical Imaging.

Research Process

a) Research Steps

Step 1: Multi-template Selection and Image Preprocessing

To eliminate diagnostic variations caused by different spatial transformations, multiple brain templates were selected for data augmentation. Specifically, each sMRI image was transformed to different template spaces via spline interpolation, and 11 templates were selected using the Affinity Propagation Clustering algorithm, including the Colin27 template and 10 templates selected from the ADNI and NACC datasets.

Step 2: Multi-template Self-supervised Learning

A simple Siamese network was constructed to randomly transform sMRI images into two template spaces (θi and θj) and extract high-level feature embeddings using an 8-layer convolutional network. These features were then decoupled through self-supervised learning and category-supervised learning.

Step 3: Weakly Supervised Meta-information Learning

To extract metadata from the features without affecting the encoder’s discriminative ability, a module consisting of InfoGAN and a decoupler was designed. InfoGAN used adversarial training to learn the metadata.

Step 4: Mutual Information Minimization

To enhance the decoupling between class-relevant features and metadata, the Contrastive Log-Ratio Upper Bound (CLUB) method was used to minimize mutual information between class-relevant features and metadata.

Step 5: Model Training

The model was trained and validated on the ADNI and NACC datasets, using the Adam optimizer with a learning rate of 0.0001 and a batch size of 6, iteratively optimized for 100 epochs.

Research Results

b) Main Results of the Study

Step 1: Results of Multi-template Self-supervised Learning

1. The multi-template selection process showed that the selected templates effectively mitigated diagnostic variability due to different spatial transformations.
2. The use of a Siamese network for self-supervised learning significantly improved the discriminative ability of feature extraction, with an average accuracy increase of approximately 5%.

Step 2: Results of Meta-information Learning and Mutual Information Minimization

1. InfoGAN was used for weakly supervised meta-information learning and reconstructing features to maintain consistency.
2. After mutual information minimization, the accuracy in AD diagnosis and MCI conversion prediction improved by about 3% and 1%, respectively.

Cross-dataset Validation

Experimental results on two multi-center datasets (ADNI and NACC) demonstrated that MMRN outperformed current state-of-the-art methods in AD diagnosis, mild cognitive impairment (MCI) conversion prediction, and in tasks distinguishing normal control (NC), MCI, and AD.

Research Conclusion and Value

c) Research Conclusion

This study proposes a novel network model for Alzheimer’s disease diagnosis using structural MRI, combining multi-template learning and meta-information regularization. Experimental results indicate that this method significantly outperforms current advanced methods in multiple tasks.

d) Research Highlights

1. Multi-template Selection significantly reduced diagnostic variability due to different spatial transformations, enhancing the reliability of feature extraction.
2. The combination of Weakly Supervised Meta-information Learning and Mutual Information Minimization enhanced the discriminative ability of class-relevant features while avoiding confounding effects from metadata.

Other Valuable Information

This research demonstrates the importance of metadata in neuroimaging analysis and shows that weak supervision and regularization techniques can significantly improve diagnostic model performance. Future research can further optimize existing methods, such as employing more data augmentation techniques, introducing more complex neural network architectures, or applying the method to the diagnosis of other types of neurological diseases to further validate and extend the proposed method.

Through this study, the authors showcase how cross-disciplinary knowledge and innovative algorithms work together to advance modern medical technology, especially in achieving significant progress in the early detection and diagnosis of complex brain diseases. This not only helps improve diagnostic accuracy but also provides crucial data support for early intervention and treatment of related diseases.