Page 121 - 南京医科大学学报自然科学版
P. 121
第44卷第2期 杜 悦,史中青,戚占如,等. 一种超声心动图关键帧智能检测方法[J].
2024年2月 南京医科大学学报(自然科学版),2024,44(2):253-262 ·259 ·
5 5
4 4
( ED ) 3 ( ES ) 3
AFD 2 AFD 2
1 1
0 0
A2C A3C A4C A2C A3C A4C
图7 ED和ES模型预测与真实标签一致性对比
Figure 7 Comparison of consistency between ED and ES model prediction and label
A ED ED ES ES
••• ••• •••
Frame No 12 13 14 37 38 39
B ED ED ES& ES
••• ••• •••
Frame No 7 8 9 29 30 31
C ED& ED ES& ES
••• ••• •••
Frame No 3 4 5 24 25 26
A:Model prediction of A2C views and examples of true labels. B:Model prediction of A3C views and examples of true labels. C:Model prediction
of A4C views and examples of true labels. ED、ES:true label;ED 、ES :model prediction
图8 A2C、A3C和A4C切面上ResNet+VST模型关键帧检测结果与真实标签对应视频帧示例
Figure 8 Examples of video frames corresponding to the detection results of key frames of ResNet+VST model and labels
on A2C,A3C and A4C views
异(P < 0.05)。Tukey检验结果进一步证明,ResNet+ 2.2 EchoNet⁃Dynamic⁃Tiny数据集
VST 模型与 3D CNN+LSTM 以及 ResNet+LSTM 模型 从 EchoNet⁃Dynamic⁃Tiny 数据集中随机挑选 1
之间均存在显著性差异(P < 0.05)。 个视频,将ResNet+VST模型的关键帧检测结果与视
表3 南京鼓楼医院数据集A4C切面不同模型ED、ES帧检测误差与推理时间对比
Table 3 Comparison of detection error and inferencing time of ED and ES frames of different models on A4C view of Nan⁃
jing Drum Tower Hospital dataset (x ± s)
Model
Performance F P
A B C
AFD(ED) 1.270 ± 1.170 *# 1.810 ± 1.690 # 1.900 ± 1.880 5.535 0.004
AFD(ES) 1.450 ± 1.380 *# 1.920 ± 1.850 # 1.650 ± 1.560 3.591 0.028
Inference time(s) 0.021 ± 0.002 *# 0.157 ± 0.009 # 0.136 ± 0.005 >100.000 <0.001
#
*
A:ResNet+VST model. B:3D CNN+LSTM model. C:ResNet+LSTM model. Compared with B,P < 0.05;Compared with C,P < 0.05.