Deep Learning Multimodal MR Image Synthesis Based on Quantitative Imaging Technology

Inhalt:
Magnetic resonance imaging (MRI) plays a vital role in clinical diagnosis. MRI images of different modalities capture diverse tissue characteristics, providing clinicians with more comprehensive pathological insight. However, in clinical practice, factors such as cost, scanning time, and patient condition often prevent the acquisition of all desired modalities, thereby compromising diagnostic precision and completeness. Generating missing modality images from available ones has thus emerged as an effective strategy to supplement imaging information without requiring additional scans. This paper introduces a generative network with robust anti-interference capabilities, designed to synthesize high-quality missing modality images. The proposed network integrates the global modeling capacity of Transformers for capturing longrange dependencies with the local feature extraction strengths of ResNet to enhance structural detail representation. It effectively leverages structural information from existing modalities to infer missing ones. Notably, the network demonstrates the ability to generate artifact-free images even when the input contains motion artifacts, maintaining quality comparable to that of real images. Experimental results confirm that the proposed method is robust to motion degradation and achieves high accuracy and visual fidelity.