Objective: To develop a new 3D multi-sequence relation attention network for exploring the complementary and correlation information of different magnetic resonance imaging (MRI) modalities and improving the diagnostic performance of HLA-B27-negative axial spondyloarthropathy (axSpA).
Methods: We retrospectively collected T1-weighted imaging (T1WI) and fat suppuration MRI (FS-MRI) data and clinical data of 375 HLA-B27-negative patients from the Third Affiliated Hospital of Southern Medical University (including 164 axSpA and 211 non-axSpA patients) and 49 patients from Nanhai Hospital (including 27 axSpA and 22 non-axSpA patients) between January, 2010 and August, 2021.A 3D relation attention network MSFANet based on multi-sequence MRI was used for automatic diagnosis of axSpA against non-axSpA in these patients.MSFANet consisted of a shallow shared feature learning module and a class-aware feature learning module, and latter module used a 3D multi-sequence relation attention mechanism to refine and fuse multi-sequence MRI features.A hybrid loss function was used to enhance the recognition ability of MSFANet by learning the loss weight coefficients of different branches to improve the classification performance.
Results: The experimental results demonstrated that MSFANet outperformed several state-of-the-art fusion algorithms (P<0.05) with AUC, accuracy, sensitivity, and specificity of 0.840, 77.93%, 83.70%, and 70.29% in the internal validation set, and of 0.783, 74.47%, 82.43% and 70.40% in the independent external validation set, respectively.The ablation studies showed that under the same architecture, the fusion model was superior to single-sequence models, which confirmed the effectiveness and necessity of fusing multi-sequence MRI.The visualization results demonstrated that MSFANet could focus on learning information from abnormal areas on MRI during the classification.
Conclusion: We successfully constructed a 3D deep neural network based on multi-sequence MRI for differential diagnosis of HLA-B27-negative axSpA against nonaxSpA and verified the effectiveness of the multisequence relation attention mechanism for promoting classification performance of the network.
目的: 建立一种新的3D多序列关系注意力网络, 通过探索不同磁共振成像(MRI)序列图像的互补和相关信息, 提升对人类白细胞抗原(HLA)-B27阴性中轴性脊柱关节病(axSpA)的诊断性能。
方法: 回顾性收集2010年1月~2021年8月南方医科大学第三附属医院(TAH组)的375例和南海医院(NHH组)的49例HLA-B27阴性参与者(TAH组: 164例axSpA, 211例非axSpA; NHH组: 27例axSpA, 22例非axSpA)的两种参数MRI, 包括T1加权图像(T1WI)和压脂序列MRI (FS-MRI), 以及相关临床数据。提出一个基于多序列MRI的3D关系注意力网络MSFANet, 实现对HLA-B27阴性axSpA与非axSpA的自动鉴别诊断。MSFANet由一个浅层共享特征模块和一个类感知特征学习模块组成, 其中类感知特征学习模块采用3D多序列关系注意力机制对多序列MRI特征进行细化和融合。提出一种混合损失函数, 通过学习不同支路的损失权重系数来提升MSFANet对序列特征的识别能力, 从而增强分类性能。
结果: 实验结果表明, MSFANet优于其它几种最先进的多序列融合算法, 其中内部验证集上的AUC、准确度、敏感度和特异度分别达到了0.840, 77.93%, 83.70%和70.29%, 独立外部验证集(NHH)上的上述性能分别达到了0.783, 74.47%, 82.43%和70.40%。各项差异均具有统计学意义(P<0.05)。此外, 消融实验显示, 相同框架下, MSFANet的性能优于基于单序列MRI的模型, 证实了融合多序列MRI的有效性和必要性。深度可视化技术显示MSFANet在分类过程中集中于学习图像异常区域的信息。
结论: 本研究成功构建基于多序列MRI的3D深度神经网络对HLA-B27阴性axSpA和非axSpA进行鉴别诊断, 并验证了采用多序列关系注意力机制对提升网络分类性能的有效性。
Keywords: 3D multi-sequence relation attention; HLA-B27 negative; axSpA diagnosis; hybrid loss; magnetic resonance imaging.