基于卷积神经网络的精子形态学分类研究

Research on sperm morphological classification based on convolutional neural network

目的:为提高精子形态学分类的准确性,提出一种基于卷积神经网络的精子分类模型。方法:使用EfficientNetB0作为基础模型,通过数据预处理增强、迁移学习以及余弦衰减进行微调,构建FT-EfficientNet模型。在精子公开数据集SCIAN-Morpho和HuSHeM上进行分类实验,利用5折交叉验证对数据集进行分割与验证,并与级联式的支持向量机(cascade ensemble of support vector machines,CE-SVM)模型、基于块的自适应字典学习(adaptive patchbased dictionary learning,APDL)模型、微调可视几何组(fine tuning of visual geometry group,FT-VGG)模型、人类精子头部形态分类(morphological classification of human sperm heads,MC-HSH)模型、迁移学习(transfer learning,TL)模型的分类结果进行对比。在SCIAN-Morpho数据集中进行消融实验,验证不同微调方法对模型的影响。结果:FT-EfficientNet模型在SCIAN-Morpho验证集上的准确率、精确度及F_(1)分数分别为64.1%、63.8%和64.8%,优于CE-SVM、APDL、FT-VGG、MC-HSH模型,召回率为65.2%,略低于MC-HSH模型(68.0%)。FT-EfficientNet模型在HuSHeM验证集上的准确率、精确度、F_(1)分数、召回率分别为95.4%、95.8%、95.4%和96.0%,略低于TL模型,但优于CE-SVM、APDL、FT-VGG、MC-HSH模型。消融实验结果表明,FT-EfficientNet模型应用的微调方法所得结果最优。结论:基于卷积神经网络的精子分类模型能够完成精子形态学分类,提升分类的准确度及性能。

Objective To propose a sperm classification model based on convolutional neural network to enhance the accuracy of sperm morphological classification.Methods A FT-EfficientNet model was constructed using EfficientNetB0 as the base model,which was fine-tuned by data preprocessing enhancement,transfer learning and cosine decay.Classification experi-ments were performed on the sperm public datasets SCIAN-Morpho and HuSHeM,and the datasets were segmented and vali-dated using 5-fold cross-validation.The classification results by the FT-EfficientNet model were compared with those by the cascade ensemble of support vector machines(CE-SVM)model,the adaptive patch-based dictionary learning(APDL)model,fine tuning of visual geometry group(FT-VGG)model,morphological classification of human sperm heads(MH-HSH)model and transfer learning(TL)model.Ablation experiments were performed in the SCIAN-Morpho dataset to verify the effect of different fine-tuning methods on the model.Results The FT-EfficientNet model proposed had the accuracy,precision and F1 score on the SCIAN-Morpho validation set being 64.1%,63.8%and 64.8%,respectively,which were better than CE-SVM,APDL,FT-VGG and MC-HSH models.The recall rate of the model proposed(65.2%)was slightly lower than that of MC-HSH model(68.0%).The accuracy,precision,F1 score and recall rate on the HuSHeM validation set was 95.4%,95.8%,95.4%and 96.0%,respectively,which were slightly lower than those of TL model while better than those of CE-SVM,APDL,FT-VGG and MC-HSH models.Ablation experiments showed the FT-EfficientNet model behaved the best in fine-tuning.Conclusion The sperm classification model based on convolutional neural network facilitates sperm morphology classification with high accuracy and performance.