Synthesizing Contrast-Enhanced T1 MR Image Using Multiparametric Sequences and Attention to Brain Tumor

ObjectiveThe administration of gadolinium-based contrast agents (GBCAs) for acquiring contrast-enhanced T1-weighted magnetic resonance imaging (T1C MRI) is associated with potential safety risks, including tissue deposition and nephrogenic systemic fibrosis. This work aims to develop a deep learning framework for synthesizing high-fidelity T1C MRI directly from multi-parametric, non-contrast sequences, thereby eliminating the need for GBCAs while preserving critical diagnostic information for brain tumor assessment. MethodsWe propose ZeroCEMR, a novel two-stage deep learning framework for GBCA-free T1C MRI synthesis. The model was developed and evaluated using multi-institutional datasets, including the BraTS 2021 benchmark and a clinical cohort. The first stage employs a Global Anatomical Encoder and YOLO-based tumor detectors to extract whole-brain structural features from T1-weighted (T1w), T2-weighted (T2w), and Fluid-Attenuated Inversion Recovery (FLAIR) images. These global and tumor-specific features are fused and decoded to generate a preliminary T1C image. The second stage refines this initial prediction by incorporating Diffusion-Weighted Imaging (DWI) and Apparent Diffusion Coefficient (ADC) sequences. Key innovations include an inter-stage residual fusion mechanism, a Squeeze-and-Excitation attention block for enhanced lesion sensitivity, and a multi-scale Feature Pyramid Network integrated within a U-Net architecture to accurately capture both macroscopic anatomy and microscopic pathological details. ResultsThe proposed ZeroCEMR framework achieved superior performance, with a peak signal-to-noise ratio (PSNR) of 37.82 and a structural similarity index measure (SSIM) of 0.98, significantly outperforming existing methods. Ablation studies confirmed the individual contribution of each architectural component, demonstrating progressive improvements in both quantitative metrics and qualitative lesion delineation. ConclusionZeroCEMR establishes a new state-of-the-art framework for synthetic T1C MRI generation. By effectively leveraging multi-parametric data and a lesion-focused architecture, our framework generates clinically viable contrast-enhanced images without GBCA administration. This approach represents a significant step towards safer, quantitative neuro-oncological imaging workflows.