Triamese-ViT: A 3D-Aware Method for Robust Brain Age Estimation from MRIs

Abstract: The integration of machine learning in medicine has significantly improved diagnostic precision, particularly in the interpretation of complex structures like the human brain. Diagnosing challenging conditions such as Alzheimer's disease has prompted the development of brain age estimation techniques. These methods often leverage three-dimensional Magnetic Resonance Imaging (MRI) scans, with recent studies emphasizing the efficacy of 3D convolutional neural networks (CNNs) like 3D ResNet. However, the untapped potential of Vision Transformers (ViTs), known for their accuracy and interpretability, persists in this domain due to limitations in their 3D versions. This paper introduces Triamese-ViT, an innovative adaptation of the ViT model for brain age estimation. Our model uniquely combines ViTs from three different orientations to capture 3D information, significantly enhancing accuracy and interpretability. Tested on a dataset of 1351 MRI scans, Triamese-ViT achieves a Mean Absolute Error (MAE) of 3.84, a 0.9 Spearman correlation coefficient with chronological age, and a -0.29 Spearman correlation coefficient between the brain age gap (BAG) and chronological age, significantly better than previous methods for brian age estimation. A key innovation of Triamese-ViT is its capacity to generate a comprehensive 3D-like attention map, synthesized from 2D attention maps of each orientation-specific ViT. This feature is particularly beneficial for in-depth brain age analysis and disease diagnosis, offering deeper insights into brain health and the mechanisms of age-related neural changes.

• The paper introduces Triamese-ViT, a novel multi-view Vision Transformer architecture for precise brain age estimation using 3D MRI data.
• It integrates three distinct ViTs processing axial, sagittal, and coronal views to synthesize a comprehensive 3D attention map.
• Evaluation on 1351 MRI scans achieved an MAE of 3.87 and a Spearman correlation of 0.93, outperforming state-of-the-art CNN and ViT models.

Triamese-ViT: A 3D-Aware Method for Robust Brain Age Estimation from MRIs

Introduction

The precise estimation of brain age plays a pivotal role in the early detection of neurological disorders, such as Alzheimer's disease. This domain has traditionally relied on convolutional neural networks (CNNs) to process 3D MRI scans. However, the emerging use of Vision Transformers (ViTs) presents a promising alternative due to their superior detail resolution and interpretability. Nevertheless, 3D applications of ViTs often fall short without particular adaptions for multidimensional data. The paper introduces Triamese-ViT, an innovative adaptation specifically tailored for brain age estimation, which integrates multiple ViTs to effectively capture 3D anatomical information.

Architecture and Methodology

The Triamese-ViT architecture capitalizes on the design of Siamese Networks, utilizing three independently functioning ViTs oriented along different MRI axis views: axial, sagittal, and coronal. Each view is processed by transforming MRI slices into patches, embedding them with positional information, and feeding them into the Transformer encoder. The novelty of Triamese-ViT arises from its comprehensive 3D-like attention map, synthesizing insights from these orientation-specific outputs through an additional Triamese MLP layer.

Figure 1: The structure of Triamese-ViT, showing the process of extracting and integrating data from three distinct viewpoints.

Performance Evaluation

The Triamese-ViT showcases impressive performance metrics on a dataset of 1351 MRI scans, achieving a Mean Absolute Error (MAE) of 3.87, a Spearman correlation of 0.93, and a Spearman correlation between brain age gap (BAG) and chronological age of -0.29. This surpasses the state-of-the-art methods, highlighting its effectiveness in accuracy and bias reduction. These metrics underscore the model's ability to reliably estimate brain age while minimizing the potential biases evident in alternative methods.

Comparison with State-of-the-Art

Comparative analysis against other prominent approaches such as ResNet, VGG, and Global-Local Transformer reveals Triamese-ViT's superior capability. With the lowest MAE and highest fairness score, the proposed model demonstrates significant advancements over traditional CNN-based and other ViT-based models, reinforcing its benefits in capturing complex 3D brain structures.

Explainability and Interpretability

The use of attention maps within Triamese-ViT greatly enhances interpretability. Occlusion sensitivity analysis, juxtaposed with the model's attention map outputs, reveals critical regions like the Basal Ganglia, Thalamus, and Midbrain as decisive elements, corresponding with known neurological implications in age-related brain diseases.

Figure 2: Comparison of attention maps and occlusion analysis illuminates the informative areas Triamese-ViT focuses on during brain age estimation.

Discussion

A key strength of Triamese-ViT lies in its synthesis of CNN and ViT advantages, combining attention to detail with comprehensive structural assessments. While CNNs may falter in capturing broad image contexts and ViTs in deep 3D interpretations, the Triamese architecture transcends these limitations by effectively unifying spatial perspectives. It demonstrates significant potential in both clinical and research contexts where transparency and detailed analysis are paramount.

Conclusion

The Triamese-ViT introduces a novel paradigm in brain age estimation through its integration of multiple view-specific Vision Transformers. Its exceptional accuracy and reduced bias promise broad applications in medical diagnostics and AI research. This foundational advancement lays the groundwork for future exploration into more complex, multi-faceted neural diagnostic tools.