Objectives: To select four algorithms with relatively balanced complexity and accuracy among deep learning image classification algorithms for automatic diatom recognition, and to explore the most suitable classification algorithm for diatom recognition to provide data reference for automatic diatom testing research in forensic medicine.
Methods: The "diatom" and "background" small sample size data set (20 000 images) of digestive fluid smear of corpse lung tissue in water were built to train, validate and test four convolutional neural network (CNN) models, including VGG16, ResNet50, InceptionV3 and Inception-ResNet-V2. The receiver operating characteristic curve (ROC) of subjects and confusion matrixes were drawn, recall rate, precision rate, specificity, accuracy rate and F1 score were calculated, and the performance of each model was systematically evaluated.
Results: The InceptionV3 model achieved much better results than the other three models with a balanced recall rate of 89.80%, a precision rate of 92.58%. The VGG16 and Inception-ResNet-V2 had similar diatom recognition performance. Although the performance of diatom recall and precision detection could not be balanced, the recognition ability was acceptable. ResNet50 had the lowest diatom recognition performance, with a recall rate of 55.35%. In terms of feature extraction, the four models all extracted the features of diatom and background and mainly focused on diatom region as the main identification basis.
Conclusions: Including the Inception-dependent model, which has stronger directivity and targeting in feature extraction of diatom. The InceptionV3 achieved the best performance on diatom identification and feature extraction compared to the other three models. The InceptionV3 is more suitable for daily forensic diatom examination.
目的: 选择深度学习图像分类算法中复杂性和准确性较为平衡的4种算法进行硅藻的自动识别，探究最适用于硅藻识别的分类算法，为法医学自动化硅藻检验研究提供数据参考。方法: 建立真实水中尸体肺组织消化液涂片的“硅藻”“背景”小样本量数据集（20 000张），用于4种算法（VGG16、ResNet50、InceptionV3和Inception-ResNet-V2）模型的训练、验证和测试。绘制受试者工作特征曲线、混淆矩阵并计算召回率、查准率、特异性、准确率及F1分数，对各模型性能进行系统性评估。结果: InceptionV3的硅藻识别性能明显优于其他3种算法，具有更为均衡的硅藻查全（89.80%）与查准（92.58%）性能；VGG16和Inception-ResNet-V2的硅藻识别性能相当，虽无法做到硅藻查全与查准的性能均衡，但其识别能力尚可接受；ResNet50的硅藻识别性能最低，其召回率仅为55.35%。在特征提取上，4种模型均提取到了硅藻和背景的特征，且都以硅藻区域为主要识别依据。结论: 包含Inception结构的模型，在硅藻特征提取方面具有更强的指向性和靶向性。其中，InceptionV3算法能够更为准确、靶向地提取到硅藻特征，具有最优的硅藻识别性能，更适合应用于日常法医学硅藻检验。.
Keywords: artificial intelligence; convolutional neural network; deep learning; diatom test; drowning; forensic pathology.