基于注意力机制和MRI的深度学习模型预测中轴型脊柱关节炎骶髂关节新骨形成进展

A self-attention-based deep learning model predicts the progression of new bone formation in the sacroiliac joints of patients with axial spondylarthritis on MRI

目的探讨基于注意力机制的深度学习(deep learning,DL)模型在骶髂关节(sacroiliac joint,SIJ)MRI冠状位T1序列图像预测中轴型脊柱关节炎(axial spondylarthritis,axSpA)患者新骨形成进展的临床价值。材料与方法回顾性分析2010年1月至2022年12月期间在南方医科大学第三附属医院诊断为axSpA的351例患者的初诊和随访1年、2年或3年的MRI图像,以8∶1∶1的比例随机分入训练集,验证集和测试集。开发基于注意力机制的Bifpn-YOLOv8模型,同时构建三个基线模型(YOLOv8、YOLOv7、Faster-RCNN)用于同Bifpn-YOLOv8比较模型效能。使用全类平均准确度(mean average precision,mAP)、F1分数、准确度、召回率、具体情境的常规物体(Common Objects in Context,COCO)指标评估各模型预测新骨形成进展的性能。其中,mAP_(50)和mAP_(50:95)分别代表不同交并比阈值下的全类平均准确度,COCO指标的平均准确度(average precision,AP)如AP,AP_(50),AP_(75)同理。结果Bifpn-YOLOv8模型在验证和测试集上均取得了良好的预测性能。同基线模型相比,该模型在测试集上取得了最优的mAP_(50)和mAP_(50:95),为83.8%和50.4%,结果同三个基线模型差异均存在统计学意义(P均<0.05)。同样,Bifpn-YOLOv8模型在测试集上获得了较基线模型更优的COCO指标,AP、AP_(50)、AP_(75)分别为50.5%、82.3%、58.6%。结论基于注意力机制的Bifpn-YOLOv8模型可以利用MRI图像有效预测axSpA患者SIJ新骨形成进展,该模型有望成为评估新骨形成进展的临床工具,辅助医师对axSpA患者进行临床决策和管理。

Objective:To investigate the clinical significance of deep learning(DL)model based on self-attention mechanisms in predicting the progression of new bone formation on coronal T1-weighted MR images of sacroiliac joints(SIJ)in patients with axial spondylarthritis(axSpA).Materials and Methods:We conducted a retrospective analysis of MRI data(with one-year,two-year or three-year follow-up duration)for 351 axSpA patients who were diagnosed at the Third Affiliated Hospital of Southern Medical University from January 2010 to December 2022.The patients were randomly allocated into training,validation,and test sets in a 8∶1∶1 ratio.The Bifpn-YOLOv8 model based on self-attention mechanisms was developed.And another three baseline models(YOLOv8,YOLOv7,Faster-RCNN)were constructed to compare model performance with Bifpn-YOLOv8.We evaluated the predictive performance of each model using metrics such as mean average precision(mAP),F1 score,accuracy,recall,and Common Objects in Context(COCO)evaluation metrics.Among them,mAP_(50) and mAP_(50:95) indicates the mean average precision at different intersection over union thresholds,respectively.The average precision(AP)of COCO metrics,such as AP,AP_(50),AP_(75),follows the same principle.Results:The Bifpn-YOLOv8 model exhibited good predictive performance on both validation and test sets.In comparison to the baseline models,Bifpn-YOLOv8 achieved the highest mAP_(50) and mAP_(50:95) on the test set,with values of 83.8%and 50.4%,respectively.The results were statistically significant(all P<0.05)compared to each baseline model.Similarly,the Bifpn-YOLOv8 model outperformed the baseline models on the test set with superior COCO evaluation metrics(AP:50.5%,AP_(50):82.3%,AP_(75):58.6%).Conclusions:The self-attention-based Bifpn-YOLOv8 model could effectively predicting the progression of new bone formation in the SIJ on MR images of axSpA patients.This model is poised to become a valuable clinical tool for evaluating the progression of new bone formation,providing assistance to physicians in clinical decision-making and management of axSpA patients.