A multi-center study of transformer-based CNNs for multiple sclerosis lesion segmentation on 3D FLAIR MRI

Area of Science:

• Medical Imaging Analysis
• Deep Learning in Neuroscience
• Computational Pathology

Background:

• Accurate segmentation of multiple sclerosis (MS) lesions on FLAIR MRI is crucial for diagnosis and monitoring.
• Existing automated methods often struggle with variability across different clinical datasets and centers.
• Deep learning models, particularly convolutional neural networks (CNNs) and Transformers, show promise for improving segmentation tasks.

Purpose of the Study:

• To develop and evaluate a novel Transformer-CNN framework for automated MS lesion segmentation on FLAIR MRI.
• To benchmark the proposed framework against established models like U-Net and DeepLabV3.
• To assess the model's segmentation accuracy and its performance across diverse clinical datasets using rigorous cross-validation.

Main Methods:

• A dataset of 1,800 3D FLAIR MRI scans from five centers was utilized with 5-fold cross-validation.
• Preprocessing involved isotropic resampling, intensity normalization, and bias field correction.
• The Transformer-CNN model integrated CNNs for local features and Transformers for global context, enhanced by data augmentation.

Main Results:

• The Transformer-CNN achieved superior performance with a Dice score of 92.3% and IoU of 91.4%.
• It demonstrated excellent across-center robustness with the highest GDSC (91.3%) and ICC (96.5%), and lowest CFDD (1.05%).
• Compared to U-Net (Dice 83.0%) and DeepLabV3 (Dice 85.1%), the Transformer-CNN showed significantly better accuracy and consistency.

Conclusions:

• The Transformer-CNN framework significantly outperforms U-Net and DeepLabV3 in MS lesion segmentation accuracy.
• The model exhibits robust generalization across diverse clinical datasets, showing minimal variability.
• This Transformer-CNN offers a practical and reliable tool for automated MS lesion segmentation and clinical monitoring.