Domain-Adaptive MRI Learning Model for Precision Diagnosis of CNS Tumors

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Diagnosing central nervous system (CNS) tumors via MRI is challenged by significant variability in scanner hardware, acquisition protocols, and intensity characteristics across clinical sites.
• These domain shifts diminish the reliability of automated diagnostic models.
• Existing methods struggle to generalize across diverse imaging datasets, limiting their real-world applicability.

Purpose of the Study:

• To develop a robust domain-adaptive learning model for accurate CNS tumor diagnosis from multi-center MRI data.
• To enhance the reliability and generalizability of automated MRI analysis in the presence of significant domain shifts.
• To improve segmentation performance and diagnostic accuracy for CNS tumors across varied imaging conditions.

Main Methods:

• A hybrid 3D Convolutional Neural Network (CNN)-transformer encoder, termed Domain-Adaptive MRI Learning Model (DA-MLM), was developed.
• The model incorporates adversarial alignment, contrastive regularization, and covariance-based feature harmonization for domain adaptation.
• Multi-sequence MRI inputs (T1, T1-contrast enhanced (T1ce), T2, and Fluid-Attenuated Inversion Recovery (FLAIR)) were processed using multi-scale convolutional layers and global self-attention.

Main Results:

• DA-MLM achieved high performance on the BraTS 2020 dataset, with 94.8% accuracy, 93.6% macro-F1, and 96.2% AUC, outperforming benchmarks by 2-4%.
• Superior segmentation Dice scores were obtained: 93.1% (whole tumor), 91.4% (tumor core), and 89.5% (enhancing tumor), surpassing existing methods by 2-3.5%.
• On the REMBRANDT dataset, DA-MLM showed 92.3% accuracy and 3-7% improved segmentation over U-Net and expert annotations, with significant robustness to noise, contrast shifts, and motion artifacts.

Conclusions:

• DA-MLM demonstrates superior accuracy, segmentation fidelity, and robustness to imaging perturbations.
• The model exhibits strong cross-domain generalization capabilities, essential for multicenter MRI applications.
• DA-MLM's performance indicates its suitability for deployment in clinical settings with unavoidable variations in MRI acquisition parameters.