基于多尺度焦点自注意力的膝骨关节炎分级方法  

Knee osteoarthritis grading method based on multi⁃scale focal self⁃attention

作　　者：蔡晓宇 汪宇玲[1] CAI Xiaoyu;WANG Yuling(College of Information Engineering,East China University of Technology,Nanchang 330000,China)

机构地区：[1]东华理工大学信息工程学院,江西南昌330000

出　　处：《现代电子技术》2025年第9期15-23,共9页Modern Electronics Technique

基　　金：国家自然科学基金项目(62066003);国家留学基金项目(CSC202208360143)。

摘　　要：膝骨关节炎(KOA)的早期检测与分级是改善慢性关节炎病患生活质量的重要诊断依据。文中提出一种基于多尺度焦点自注意力的KOA分级模型MSFFormer。该模型将显性空间先验技术与自注意力机制相结合,设计Y轴向空间衰减机制来构建二维显性空间先验矩阵,并结合自注意力机制提出多尺度焦点注意力使模型更加关注KOA图像病变区域与相邻区域的关联特征,同时减少无关背景产生的注意力冗余。在公开膝骨关节炎数据集OAI上与其他先进方法相比,MSFFormer在早期KOA二分级任务与KOA 5类KL分级任务中Accuracy指标分别提升了1.73%、0.88%,证明所提方法更利于KOA早期检测与分级的辅助诊断。Early detection and grading of knee osteoarthritis(KOA)is an important diagnostic basis for improving the quality of life of chronic arthritis patients.Therefore,a KOA hierarchical model MSFFormer(multi-scale focal transformer)based on multi-scale focal self-attention is proposed.In the model,the dominant spatial prior technology is combined with the self-attention mechanism,the Y-axial spatial decay mechanism is designed to construct the two-dimensional dominant spatial prior matrix,and the multi-scale focus attention is proposed in combination with the self-attention mechanism to make the model pay more attention to the correlation features between the lesion area and adjacent areas in KOA images,while reducing the attention redundancy caused by irrelevant background.In comparison with the other advanced methods in the open knee osteoarthritis dataset OAI(osteoarthritis initiative),the accuracy of MSFFormer in the early KOA two-grading tasks and KOA five-class KL(Kellgren and Lawrence system)grading tasks has improved by 1.73%and 0.88%,respectively.Therefore,it is verified that the MSFFormer is more conducive to the early detection and grading of KOA.

关 键 词：计算机辅助诊断 X线片图像 膝骨关节炎 深度学习 TRANSFORMER 自注意力机制 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程]