分子影像学杂志2025，Vol.48Issue(11)：1358-1363,6.DOI:10.12122/j.issn.1674-4500.2025.11.06
超分辨深度学习模型优化小肝癌磁共振扩散加权成像质量
Deep learning-based super-resolution optimization enhances diffusion-weighted MRI quality for small hepatocellular carcinoma
摘要
Abstract
Objective To develop a deep learning-based super-resolution reconstruction framework for small hepatocellular carcinoma(sHCC)in magnetic resonance imaging,designed to synergistically enhance lesion conspicuity and diagnostic utility by jointly optimizing high-frequency detail preservation and anatomical fidelity in thin-slice diffusion-weighted imaging(DWI).Methods A retrospective study was conducted on 3-mm DWI data from 300 patients with sHCC admitted from December 2022 to June 2024.The dataset was randomly divided into training and test sets at an 8:2 ratio.A dual-branch super-resolution model was constructed,comprising a content branch for extracting global features through cascaded gradient Transformer blocks and a gradient branch for enhancing structural information via gradient Transformer blocks.The model incorporated two novel components:(1)a cross local-enhanced self-attention module to optimize interactions between pixel-level features and global context,and(2)a channel-spatial joint attention layer that dynamically allocates weights to improve the visibility of key anatomical structures.Three senior radiologists blindly evaluated both original(OR-DWI)and super-resolved(SR-DWI)images using a 5-point Likert scale.Objective metrics,including the peak signal-to-noise ratio(PSNR)and structural similarity index(SSIM),were also calculated.Results SRDWI demonstrated significantly superior subjective ratings compared to ORDWI across multiple evaluation metrics(P<0.001),including hepatic nodule signal specificity(3.41±0.53 vs 2.47±0.50),normal liver parenchyma homogeneity(3.29±0.47 vs 2.78±0.42),artifact interference severity(2.56±0.52 vs 2.47±0.48),and overall image quality(3.15±0.49 vs 2.67±0.48).Although post-super-resolution processing resulted in a marginally increased image noise level(3.98±0.61 vs 2.87±0.46,P<0.05),the values remained well within clinically acceptable limits without compromising diagnostic interpretation of key features.Quantitative analysis further confirmed SRDWI's excellent performance,achieving a peak signal-to-noise ratio of 34.65489 dB and structural similarity index of 0.90365.Conclusion The proposed deep learning framework effectively bridges the diagnostic performance gap between thin-slice(3 mm)and conventional thick-slice DWI protocols,achieving comparable lesion characterization accuracy while mitigating motion artifacts.This advancement establishes a robust technical foundation for precision diagnosis of early-stage sHCC,where subtle anatomical details critically influence therapeutic decision-making.关键词
超分辨技术/深度学习/薄层DWI/小肝癌/磁共振成像Key words
super-resolution/deep learning/thin-slice DWI/small hepatocellular carcinoma/magnetic resonance imaging引用本文复制引用
刘旭红,刘娜红,张乾营,丁碧娇,黄莹,黄德天,何桂凤,邓娜,韩晓兵,林雅萍..超分辨深度学习模型优化小肝癌磁共振扩散加权成像质量[J].分子影像学杂志,2025,48(11):1358-1363,6.基金项目
福建省科技创新联合资金项目计划(2024Y9455) （2024Y9455）
泉州市科技计划项目(2024NY057) （2024NY057）
第910医院院级科研基金项目(910YK202307) （910YK202307）
