Objectives: To construct a YOLOv3-based model for diatom identification in scanning electron microscope images, explore the application performance in practical cases and discuss the advantages of this model.
Methods: A total of 25 000 scanning electron microscopy images were collected at 1 500× as an initial image set, and input into the YOLOv3 network to train the identification model after experts' annotation and image processing. Diatom scanning electron microscopy images of lung, liver and kidney tissues taken from 8 drowning cases were identified by this model under the threshold of 0.4, 0.6 and 0.8 respectively, and were also identified by experts manually. The application performance of this model was evaluated through the recognition speed, recall rate and precision rate.
Results: The mean average precision of the model in the validation set and test set was 94.8% and 94.3%, respectively, and the average recall rate was 81.2% and 81.5%, respectively. The recognition speed of the model is more than 9 times faster than that of manual recognition. Under the threshold of 0.4, the mean recall rate and precision rate of diatoms in lung tissues were 89.6% and 87.8%, respectively. The overall recall rate in liver and kidney tissues was 100% and the precision rate was less than 5%. As the threshold increased, the recall rate in all tissues decreased and the precision rate increased. The F1 score of the model in lung tissues decreased with the increase of threshold, while the F1 score in liver and kidney tissues with the increase of threshold.
Conclusions: The YOLOv3-based diatom electron microscope images automatic identification model works at a rapid speed and shows high recall rates in all tissues and high precision rates in lung tissues under an appropriate threshold. The identification model greatly reduces the workload of manual recognition, and has a good application prospect.
目的: 基于YOLOv3模型构建硅藻电子显微镜图像识别模型，测试在实际案例中的应用效果，探讨该模型在法医学硅藻鉴定中的优势。方法: 选取25 000张1 500×的硅藻电子显微镜图像作为初始图像集，经专家标注和图像处理后，输入YOLOv3网络训练识别模型。收集8例溺死案例肺、肝、肾组织样本的硅藻电子显微镜图像，分别在0.4、0.6、0.8的阈值下利用该模型识别硅藻，同时进行人工识别，以识别速度、召回率和查准率评价该模型的实际应用效果。结果: 模型在验证集和测试集上的平均精度均值分别为94.8%和94.3%，平均召回率分别为81.2%和81.5%。实际应用中，该模型识别速度是人工识别速度的9倍以上。阈值为0.4时，肺组织硅藻总体召回率均值达89.6%、总体查准率均值达87.8%；肝、肾组织硅藻总体召回率达100%、查准率低于5%。随着阈值的增加，各组织硅藻召回率下降、查准率上升；模型在肺组织的F1分数随着阈值升高而降低，在肝、肾组织的F1分数随着阈值升高而升高。结论: 基于YOLOv3的硅藻电子显微镜图像识别模型能够快速识别硅藻图像，并能在一定的阈值下取得良好的硅藻召回率和肺组织较高的硅藻查准率，显著减少了人工识别的工作量，具有良好的应用前景。.
Keywords: YOLOv3; artificial intelligence; deep learning; diatom test; forensic pathology; image recognition; scanning electron microscope.