文档介绍：中国医学装备大会暨 2019 医学装备展览会论文汇编
[7] 李 志 勇 , 李 鹏 伟 , 高 小 燕 , 等 . 人 工 智 能 医 学 技 术 发 展 的 聚 焦 领 域 与 趋 age preprocessing method and the deep learning object detection
model，we studied for the end-to-end model for quickly detecting plaques and plaque locations of carotid
ultrasound images generated by multiple manufacturers. Methods: Carotid plaques were marked by senior
physicians. The training set was 2，592 images of 613 patients and the test set was 497 images of 82 patients.
Firstly，based on the automatic image preprocessing method，high-quality samples with only the effective area are
generated. The training set images are trained by the Faster RCNN based on VGG16 and ResNet101 and the
Yolo-v3 based on Darknet. Through parameters optimization and train iteration，we obtained the better model for
each network. Finally，we evaluate the performance of each model in the test set based on image IoU (predicted
area and real area intersection ratio) and accuracy for each subject. Results: As for the comparison，the Faster
RCNN model based on Resnet101 had the best plaque recognition performance，with % sensitivity，%
specificity，and % accuracy. Conclusion: The results show that the end-to-end model of carotid plaque area