基于自适应序列罚权深度神经网络的膝骨关节炎等级评分

Grading scoring of knee osteoarthritis based on adaptive ordinal penalty weighted deep neural networks

膝骨关节炎(OA)是老年人活动受限和身体残疾的主要原因之一,对膝骨关节炎的早期发现和干预可以帮助病人减缓OA的恶化。目前膝骨关节炎的早期发现通过X光片进行诊断,参照Kellgren-Lawrence(KL)标准进行评分,但医师的评分相对主观,不同医生存在差异。膝骨关节炎的等级分类是个有序分类问题,序列罚权损失函数将距离真实类别越远的等级赋予了更高的罚权,因此它更适合于膝骨关节炎的等级分类。然而,已有工作中的罚权一旦给定,就不再变化,导致其训练模型常常达不到期望的结果。本文针对序列罚权损失的不足,提出一种自适应序列罚权调整策略,通过对每一个阶段(epoch)得到的混淆矩阵,反向指导惩罚权重进行微调,使得罚权矩阵能够自适应调整。进一步地,本文利用来自骨关节炎倡议组织(OAI)的X射线图像数据,在ResNet,VGG,DenseNet以及Inception等几种经典的CNN模型上验证该方法的性能。实验结果表明在膝骨关节炎KL分级任务上,本文提出的自适应序列罚权调整策略在初始罚权分差较小时,能够有效地提升模型分类精度(AC)与平均绝对误差(MAE)。

Knee osteoarthritis(OA)is one of the main causes of activity limitation and physical disability in the elderly.Early diagnosis and intervention of knee osteoarthritis can help patients slow down the deterioration of OA.At present,the early diagnosis of knee osteoarthritis is detected by X-rays and scored according to the Kellgren-Lawrence(KL)grade.However,doctors′scores are relatively subjective and vary from doctor to doctor.Grade classification of knee osteoarthritis is a matter of orderly classification.The ordinal penalty loss function assigns higher penalty weights to the classes that are further away from the ground truth,which is more suitable for knee osteoarthritis classification.In existing works,the penalty weights no longer change during training procedure,so the training model often fails to reach the expected results.In this paper,an adaptive ordinal penalty adjustment strategy is proposed to address the shortcomings of the ordinal penalty loss,in which the penalty weights are automatically tuned in reverse according to the confusion matrix obtained at each stage(epoch).Furthermore,the performance of the proposed method is validated on several classical CNN models such as ResNet,VGG,DenseNet and Inception by X-ray image data from Osteoarthritis Initiative(OAI).Experimental results show that the adaptive ordinal penalty adjustment strategy proposed in this paper can effectively improve the classification accuracy(AC)and mean absolute error(MAE)of the model when the initial weight score difference is small.