Semantic Prior Embedding for Language-Guided Medical Image Reconstruction Quality Enhancement
Year: 2026
Type: conference-paper
Abstract: Medical image reconstruction from degraded acquisitions remains a fundamental challenge in clinical imaging. This paper investigates the integration of semantic prior embeddings derived from multimodal large language models to enhance reconstruction quality in low-dose computed tomography and accelerated magnetic resonance imaging. A language-guided framework is proposed that extracts anatomical and pathological semantic features from degraded inputs and transforms them into structured prior embeddings to guide reconstruction. The semantic embedding module employs cross-attention mechanisms to fuse textual descriptions with visual features, thereby providing high-level constraints during reconstruction. Comprehensive experiments on public datasets demonstrate that the proposed approach achieves superior performance compared with strong diffusion- and score-based baselines, with peak signal-to-noise ratio improvements of 1.54 dB and 1.58 dB on low-dose computed tomography and accelerated magnetic resonance imaging benchmarks, respectively, and structural similarity gains of 0.025 and 0.024. Clinical relevance assessments confirm enhanced preservation of diagnostically significant structures and reduced artifact propagation in reconstructed images.
Author: Yali Zhang
Institution: Rice University
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203998002