SF2Former: Amyotrophic Lateral Sclerosis Identification From Multi-center MRI Data Using Spatial and Frequency Fusion Transformer

Abstract: Amyotrophic Lateral Sclerosis (ALS) is a complex neurodegenerative disorder involving motor neuron degeneration. Significant research has begun to establish brain magnetic resonance imaging (MRI) as a potential biomarker to diagnose and monitor the state of the disease. Deep learning has turned into a prominent class of machine learning programs in computer vision and has been successfully employed to solve diverse medical image analysis tasks. However, deep learning-based methods applied to neuroimaging have not achieved superior performance in ALS patients classification from healthy controls due to having insignificant structural changes correlated with pathological features. Therefore, the critical challenge in deep models is to determine useful discriminative features with limited training data. By exploiting the long-range relationship of image features, this study introduces a framework named SF2Former that leverages vision transformer architecture's power to distinguish the ALS subjects from the control group. To further improve the network's performance, spatial and frequency domain information are combined because MRI scans are captured in the frequency domain before being converted to the spatial domain. The proposed framework is trained with a set of consecutive coronal 2D slices, which uses the pre-trained weights on ImageNet by leveraging transfer learning. Finally, a majority voting scheme has been employed to those coronal slices of a particular subject to produce the final classification decision. Our proposed architecture has been thoroughly assessed with multi-modal neuroimaging data using two well-organized versions of the Canadian ALS Neuroimaging Consortium (CALSNIC) multi-center datasets. The experimental results demonstrate the superiority of our proposed strategy in terms of classification accuracy compared with several popular deep learning-based techniques.

• The paper introduces SF2Former, a dual-branch transformer model that fuses spatial and frequency features from MRI for accurate ALS identification.
• It leverages Vision Transformers for spatial encoding and the Global Filter Network for frequency analysis, outperforming traditional CNNs.
• The framework employs a majority voting scheme and integrates multi-center MRI modalities to enhance robustness in ALS diagnosis.

SF2Former: Amyotrophic Lateral Sclerosis Identification From Multi-center MRI Data Using Spatial and Frequency Fusion Transformer

Introduction

The study introduces SF2Former, a novel framework designed to differentiate patients with Amyotrophic Lateral Sclerosis (ALS) from healthy controls using MRI data from multiple centers. The research leverages the capabilities of Vision Transformers (ViTs) to distinguish subtle neurodegenerative changes in ALS, often poorly captured by traditional convolutional networks due to insignificant structural variations in neuroimaging data. By integrating features from both spatial and frequency domains—where MRI captures raw data before conversion—the framework capitalizes on the advantages of frequency domain data representation. This section summarizes the notable contributions of SF2Former, including its effective capturing of discriminative features and the implementation of a majority voting scheme for enhanced classification decisions.

Methodology

The SF2Former architecture combines spatial and frequency domain features using a dual-branch approach (Figure 1). The left branch encodes spatial features employing ViT, known for its capability to model long-range dependencies through self-attention mechanisms. Simultaneously, the right branch captures frequency domain interactions using the Global Filter Network (GFNet), applying Fourier transforms to better exploit inherent MRI data properties.

After preprocessing using tools like FreeSurfer and FSL, coronal slices are selected based on empirical analysis. The methodology also incorporates a linear fusion module to integrate spatial and frequency representations, thus refining the aggregated signal for final classification. Majority voting across coronal slices ensures robustness against false positives, effectively accommodating slice variation from single subject scans.

Figure 1: Proposed $SF^2Former$ architecture. The left branch encodes features from the spatial domain, whereas the right branch encodes features from the frequency domain. The linear fusion module assembles the classification decision for each 2D slice.

Experimental Results

Experimental validation employed multi-modal datasets from CALSNIC, focusing on three MRI modalities: T1-weighted, R2*, and FLAIR. The system demonstrated superior performance, exhibiting improved classification accuracy over existing CNN-based approaches (Table 1). Key metrics included Accuracy (ACC), Sensitivity (SEN), and Specificity (SPE), all indicating SF2Former's robustness in multi-modal, multi-center contexts.

Figure 2: The overall workflow of the proposed stages.

Figure 3: Subject-level split process for the data to train our proposed model.

Ablation and Comparative Analysis

The ablation study underscores the contribution of each architectural and methodological component to the overall performance. Removal of data normalization, augmentation, or transfer learning significantly diminished accuracy. Furthermore, comparative analysis revealed SF2Former’s distinct advantage over 2D and 3D CNN architectures, as well as previous texture-based methods, particularly with T1-weighted images.

Discussion and Implications

SF2Former's deployment of multi-domain feature extraction positions it as a notable advancement in ALS imaging biomarker research. The dual-branch transformer approach is especially appropriate for handling the multi-slice nature of MRI data while remaining robust across various imaging modalities and scanner types. The integrated majority voting scheme reliably accounts for slice variability inherent in scanned subjects.

The implications of this research are significant, fostering the future utility of MRI as a diagnostic biomarker for neurodegenerative disorders. Prospective work could further incorporate clinical and functional imaging data, enhancing diagnostic accuracy and generalization across neurodegenerative conditions.

Conclusion

SF2Former exemplifies a sophisticated approach to ALS classification using MRI, effectively overcoming limitations of previous CNN-centric models. This framework’s adaptability to different imaging modalities and its robustness in multi-center datasets make it a promising tool for ALS diagnosis and progression monitoring. Future studies could extend its applicability to integrate alternative neuroimaging and clinical data, offering expansive avenues for diagnostic exploration.