Detection of mild cognitive impairment based on attention mechanism and parallel dilated convolution

Tao Wang, Zenghui Ding, Xianjun Yang, Yanyan Chen, Yu Liu, Xiaoming Kong, Yining SunPeerJ Comput. Sci 2024 · DOI

TL;DR

This paper introduces a novel deep learning approach for detecting mild cognitive impairment (MCI) from brain MRI scans. By combining an attention gate mechanism to focus on relevant brain regions and parallel dilated convolutions to capture multi-scale contextual information, the method achieves an accuracy of 81.63% on the ADNI dataset, outperforming baseline models. This advancement holds promise for earlier and more accurate MCI diagnosis, potentially leading to better patient outcomes.

What Problem Does This Paper Address?

Mild cognitive impairment (MCI) is a critical precursor to neurodegenerative diseases like Alzheimer's disease (AD). Early detection and intervention are crucial for slowing disease progression. However, MCI is characterized by subtle changes in brain MRI scans, making them difficult to detect visually. Furthermore, MRI data often contains a lot of redundant information, which can hinder the performance of traditional deep learning methods. Existing MCI detection methods struggle to effectively pinpoint these subtle, localized abnormalities within the complex brain structure, leading to suboptimal diagnostic accuracy.

Methods

The proposed method integrates two key deep learning components into a ResNet18 backbone:

  • Attention Gate Mechanism (AG): This mechanism is designed to mimic human selective attention. It learns to highlight salient features in the lesion areas of the MRI scans while suppressing irrelevant background information. This is particularly useful for medical images where the target region (lesion) is small and localized. The AG uses a global feature vector to guide the attention to local feature vectors, effectively focusing on areas of interest.
  • Parallel Dilated Convolution (Atrous Spatial Pyramid Pooling - ASPP): To address the challenge of small lesion sizes and the need for contextual information, the authors employ parallel dilated convolutions. Unlike standard convolutions that can lose spatial information through pooling, dilated convolutions expand the receptive field without downsampling. The ASPP module uses multiple dilated convolutions with different dilation rates in parallel. This allows the network to capture contextual information at various scales, providing a richer representation of the lesion area and improving the ability to recognize small targets.

These components are integrated into an improved ResNet18 architecture. The deepest feature map from the ResNet blocks is used as a gating signal for the attention mechanism, providing contextual information to earlier layers. The outputs from the ASPP module are then fused with this gated information to create a comprehensive feature representation.

Datasets

The study utilizes the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Specifically, they selected 3.0T MRI images from the ADNI database. The dataset includes:

  • 127 Normal Control (NC) subjects with 265 MRI scans.
  • 75 Mild Cognitive Impairment (MCI) patients with 265 MRI scans.

The data was split into a 75% training set and a 25% test set, with a focus on avoiding data leakage by splitting based on subjects.

Key Insights and Ideas

  • Attention for Medical Images: The paper highlights that standard attention mechanisms, while effective for natural images, might not be optimal for medical imaging due to the localized nature of lesions. The proposed Attention Gate mechanism is specifically tailored to address this by focusing on relevant regions and suppressing noise.
  • Dilated Convolutions for Small Targets: The challenge of detecting small, subtle lesions in MRI is a significant hurdle. The use of parallel dilated convolutions (ASPP) is a clever way to increase the receptive field and capture contextual information without losing crucial fine-grained details, which is essential for identifying these small abnormalities.
  • Synergy of Attention and Multi-Scale Features: The combination of attention to focus on "where" to look and dilated convolutions to understand "what" is in the context of different scales appears to be a powerful strategy for improving MCI detection accuracy.
  • ResNet18 as a Lightweight Backbone: The study demonstrates that ResNet18, despite being a relatively smaller architecture compared to deeper ResNets, offers a good balance of performance and computational efficiency, making it suitable for resource-constrained environments.

Key Results

  • The proposed method achieved an accuracy of 81.63% on the ADNI dataset for MCI detection.
  • This represents an improvement of approximately 6.8% compared to the basic ResNet18 model.
  • The confusion matrix shows that out of 147 test samples, 120 were correctly predicted.
  • The ROC curve analysis yielded an AUC of 83.06%.
  • Ablation studies showed that adding the ASPP module increased accuracy by about 4.76%, and integrating the Attention Gate mechanism also contributed significantly to performance improvements.
  • The method demonstrated robustness across different dataset distributions, maintaining good performance with a four-fold cross-validation approach.

Limitations and Future Work

  • Focus on Structural MRI: The current study primarily uses structural MRI data. The authors suggest that future research could explore other imaging modalities like functional MRI (fMRI), electroencephalography (EEG), and PET scans, as these can provide complementary perspectives and potentially improve diagnostic accuracy.
  • Clinical Validation: While experimental validation on public datasets shows promising results, the model needs further validation and application in real clinical settings. Real-world clinical environments present additional challenges such as data variability from different devices, missing data, and other complexities that need to be addressed.
  • Dataset Size and Variability: The authors acknowledge that the relatively small dataset size and inter-subject variability contribute to a large standard deviation in performance metrics, suggesting that larger and more diverse datasets would be beneficial.

Summary updated

BibTeX
@article{wang2024detection,
  author = {Tao Wang and Zenghui Ding and Xianjun Yang and Yanyan Chen and Yu Liu and Xiaoming Kong and Yining Sun},
  title = {{Detection of mild cognitive impairment based on attention mechanism and parallel dilated convolution}},
  journal = {PeerJ Comput. Sci.},
  year = {2024},
  doi = {10.7717/peerj-cs.2056},
}
Figure