超声心动图心尖四腔心切面图像质量智能评分研究

Intelligent scoring of quality of echocardiography images of apical four-chamber view

目的探讨基于深度学习的多尺度图像质量评价Transformer网络(MUSIQ)对超声心动图质量智能评分的可行性。方法选取2021年12月至2022年4月在四川省人民医院完成超声心动图检查的457例受检者的心尖四腔心切面动态图像(共70181帧)。由2位高年资专科医师根据图像切面质量,将图像标注为0分(极差)~5(优秀)分,所有图像按照8∶1∶1的方式划分为训练集、验证集和测试集。然后采用MUSIQ模型进行网络训练和验证,并选取在验证集上综合性能表现最好的模型作为最终模型。将测试集中MUSIQ模型预测分数与专业医师所标注的分数进行比较,以精确率、召回率、F1分数对模型的诊断性能进行评价。结果模型预测结果在0~5分的精确率平均为0.941,召回率平均为0.941;F1分数平均为0.941,均处于较高值范围。同时,当图像大小为600×800时模型在GPU上的单帧推理时间为18 ms,满足实时部署要求。结论基于深度学习的MUSIQ网络对心尖四腔心切面图像质量智能评分结果与超声医师人工打分结果一致性较高,具有较高的可行性;同时由于该方法对心尖四腔心切面没有做任何前置假设,原理上可以推广到任何标准超声切面,有利于实现超声图像质量智能化评估的标准化和规范化。

Objective To explore the feasibility of echocardiographic image quality evaluation by deep learning based Multi-scale Image Quality Evaluation Transformer network(MUSIQ).Methods We selected 70181 echocardiographic apical four-chamber images of 457 patients who underwent echocardiographic examination at Sichuan Provincial People's Hospital from December 2021 to April 2022.According to the quality of the image section,two senior specialists scored the images from 0(very poor)to 5(excellent).The images were divided into a training set,a verification set,and a test set in a ratio of 8∶1∶1.Then,we used the MUSIQ models for network training and verification,and selected the model with the best comprehensive performance in the verification set.The prediction scores of the MUSIQ model in the test set were compared with the scores given by professional doctors,and the diagnostic performance of the model was then evaluated in terms of precision,recall,and F1 scores.Results The average precision,recall,and F1 scores of the model in predictions for scores between 0 and 5 were 0.941,0.941,and 0.941,respectively,all of which were all at a high level.The inference time of the model on GPU was 18 ms for single frame images with a size of 600×800,which meets the requirement of real-time processing.Conclusion The intelligent assessment of apical four-chamber cardiac image quality by MUSIQ based on deep learning has a high degree of consistency with ultrasonographers'manual scoring results and is highly feasible.As the method does not make any antecedent assumptions for apical four-chamber cardiac sections,it can be extended to any standard ultrasound section in principle,which is conducive to the standardization of intelligent assessment of ultrasound image quality.