| 摘要: |
| 目的 针对肝囊型包虫病超声图像中小病灶难以被检测、影响疾病及时诊断与治疗的问题，提出了一种以YOLOv7为基底、用于检测肝囊型包虫病五类分型超声图像中的小病灶的方法。方法 首先，用硬件感知神经网络EfficientRep替换原特征提取主干，实现在保证精度和速度不受影响的前提下，提高对硬件设备的适配度；其次，用更优的WIOU（Wise-IOU）替换CIOU，改善了YOLOv7网络的评价指标CIOU（Complete Intersection over Union）在作为损失函数时，梯度计算效果差，导致检测精度下降的问题；最后，在主干的最后第四层加入CBAM注意力，进一步提高了模型检测精度。结果 本实验在自建的肝囊型包虫病超声图像小病灶数据集上进行了训练。结果显示，改进后的模型mAP@0.50检测精度为88.1%，相比原始的模型性能得到了提升，并超过了对比的其余主流检测方法。结论 实验说明模型能更高效的检测并分类肝囊型包虫病超声图像中的病灶位置和类别、满足临床实时性要求，为后期研究提供新思路。 |
| 关键词: 肝囊型包虫病 深度学习 目标检测 YOLOv7 EfficientRep Wise-IoU CBAM |
| DOI： |
| 投稿时间：2023-07-25修订日期：2023-09-21 |
| 基金项目:新疆高发病肝包虫疾病计算机辅助诊断方法的研究(81560294);省部共建中亚高发病成因与防治国家重点实验室(SKL-HIDCA-2020-YG);肝包虫影像人工智能判读虚拟仿真实验平台的构建研究(202110760006) |
|
| Research on Deep Learning Based Detection Method for Small Lesions in Ultrasound Images of Hepatic Cystic Hydatidosis |
| Miwueryiti·HAILATI,Renaguli·AIHEMAITINIYAZI,Kadiliya·KUERBAN,YAN Chuan bo |
| () |
| Abstract: |
| objective A method based on YOLOv7 is proposed to detect small lesions in liver cystic echinococcosis ultrasound images, which are difficult to detect and affect the timely diagnosis and treatment of the disease.method Firstly, replace the original feature extraction backbone with a hardware aware neural network EfficientRep to improve the adaptability to hardware devices while ensuring accuracy and speed are not affected; Secondly, replacing CIOU with a better WIOU (Wise IOU) improves the evaluation index CIOU (Complete Intersection over Union) of YOLOv7 network, which has poor gradient calculation performance as a loss function and leads to a decrease in detection accuracy; Finally, adding CBAM attention to the final fourth layer of the backbone further improves the model detection accuracy.result This experiment was trained on a self built dataset of small lesion ultrasound images of hepatic cystic echinococcosis. The results show that the improved model mAP @ 0.50 has a detection accuracy of 88.1%, which improves the performance compared to the original model and surpasses other mainstream detection methods compared.conclusion The experiment shows that the model can more efficiently detect and classify the location and category of lesions in ultrasound images of hepatic cystic echinococcosis, meeting clinical real-time requirements, and providing new ideas for future research. |
| Key words: hepatic cystic echinococcosis deep learning object detection YOLOv7 EfficientRep Wise-IoU CBAM |
