基于迁移学习的IVF胚胎分类方法

IVF embryo image classification method based on transfer learning

针对传统体外受精(IVF)胚胎存活力的评估主要是基于胚胎学家的主观视觉分析,但受限于观察者之间的差异,并且是一项耗时的任务.在这项研究中,将深度学习与模型迁移结合使用,开发用于胚胎评估的自动分类方法.研究包括kaggle比赛数据库中的胚胎数据集,使用在ImageNet数据集上预训练网络作为基础网络,在修改基础模型全连接分类层的基础上,将高层卷积模块的权重设置为可训练以进行微调.实验结果表明,通过微调训练获得的网络准确率高达96.93%,相比基础模型表现有显著提升,证明在计算资源有限的前提下,使用微调策略也能获得不错功能的卷积神经网络模型.研究集成了深度学习方法、延时显微镜系统和IVF电子病历平台,可实现用于胚胎评估的全自动无创系统.

The issue that traditional viability evaluation of embryo for in vitro fertilization(IVF) is primarily decided by the subjective visual observation of embryologists, but it is time-consuming and limited by the differences between the observers and themselves. In this study, we combined deep learning with model transferring to develop an automatic classification method for embryo evaluation. This research includes the embryo data set from the kaggle competition database, using the pre-trained network on the ImageNet data set as the basic model. After changing the fully connected layer of the basic model, the parameters of the high-level convolution module are set to be trainable in fine-tuning phase. In the test results, the accuracy of the network obtained by fine-tuning training is as high as 96.93%, which is significantly better than the performance of the basic model. It proves that a convolutional neural network model with good functions can be obtained by using the fine-tuning strategy under the premise of limited computing resources. This research integrates deep learning methods, time-lapse microscope machine and IVF electronic medical history. They can achieve a fully automatic non-invasive embryo assessment equipment.