基于多模态数据融合的MR图像超分辨重建  

Multi-contrast Data Fusion Network for MR Image Super-resolution

作　　者：胡彦婷[1] 樊孝喜[1] 陈建军[1] HU Yanting;FAN Xiaoxi;CHEN Jianjun(School of Medical Engineering and Technology,Xinjiang Medical University,Urumqi 830011,China)

机构地区：[1]新疆医科大学医学工程技术学院,新疆乌鲁木齐830011

出　　处：《北华大学学报（自然科学版）》2023年第6期810-819,共10页Journal of Beihua University（Natural Science）

基　　金：新疆维吾尔自治区自然科学基金项目(2020D01C157,2021D01C295);国家自然科学基金项目(62061047,62265016)。

摘　　要：核磁共振(Magnetic Resonance,MR)成像是临床医学诊断的主要影像工具之一,但受成像条件、扫描时间等因素限制,MR成像设备获取的影像存在分辨率低的问题.超分辨率重建(Super-Resolution,SR)技术是提高MR影像分辨率经济、高效的手段.然而,多数MR图像超分辨重建方法或者忽略了MR多参数成像特点,仅利用其单一模态数据进行重建,导致重建效果欠佳;或者仅是简单地在模型输入端进行多模态数据融合,未充分利用多模态数据在多层级特征空间的关联性,从而限制了SR模型重建性能的提升.为更高效地利用多模态信息提高MR影像分辨率,构建了基于多模态数据融合的MR图像超分辨重建网络(Multi-contrast Data Fusion Network for MRI Super-Resolution,MDFN).结合注意力机制和残差结构,构建特征融合模块(Feature Fusion Module,FFM),实现对多模态数据融合后特征的重新校准,有效增强高价值信息的重建作用;同时,通过级联的残差块和特征融合模块(FFM),构建特征融合残差组(Residual Group with Feature Fusion,RGFF),并通过级联RGFF模块,实现多模态特征在多层级特征空间的有效融合.在MR数据测试集上的综合试验结果表明,所构建的模型能以低的模型参数量和计算复杂度获得在主观质量和客观评价上都优于其他方法的重建结果.Magnetic Resonance(MR) imaging is one of the main tools for clinical diagnosis.However,limited by imaging conditions and scanning time,MR imaging systems always generate low-resolution image sequences.The technology of super-resolution(SR) is a cost-effective means to improve resolution of MR images.Most MR methods for MR images either ignore the multi-contrast nature of MR imaging and only utilize a single contrast data for super-resolving,or simply fuse multi-contrast MR images as the input of the model without sufficiently exploiting the correlation between multi-contrast data in multi-level feature spaces,which limit the MR image super-resolution performance.We propose Multi-contrast Data Fusion Network(MDFN) for MR image super-resolution to efficiently improve MR image resolution based on multi-contrast MR data.Specifically,we construct Feature Fusion Module(FFM) by adopting the attention mechanism and residual structure,which could recalibrate the fused multi-contrast features and enhance high contribution information.Meanwhile,we build a residual group with feature fusion(RGFF)by cascading the residual blocks and FFM,and cascade multiple RGFFs to realize effective multi-stage fusion of multi-contrast features in multi-level layers.Comprehensive evaluations on MR image dataset demonstrate that MDFN model with fewer parameters and lower computational complexity achieves better performances in terms of quantitative and qualitative measurements.

关 键 词：超分辨重建 MR多模态成像 注意力机制 特征融合 卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]