基于深度学习与医学先验知识的超声心动图切片识别

摘要

针对传统机器学习方法在围术期食管超声心动图(TEE)上进行切片识别时识别精度不够高和模型不能端到端的问题,提出了一种基于深度学习与医学先验的端到端切片识别方法.首先,提取TEE超声切片上的切片角度信息训练一个小型卷积神经网络(CNN)进行角度分类,获取分类结果即医学先验概率;然后,对整个TEE超声切片成像区域训练一个大型深度学习网络模型进行切片分类,获取验证前分类结果,即条件概率;最后通过贝叶斯方法校验获取最终的识别结果.实验结果表明,与传统方法相比结合深度学习与医学先验的切片识别方法极大地提高了TEE切片识别的精度.%The traditional machine learning methods for TEE (TransEsophageal Echocardiography) recognition have the problems of low accuracy and are not end-to-end methods.A new end-to-end method based on the deep learning and medical priori knowledge was introduced.Firsdy,a small deep CNN (Convolutional Neural Network) was used to classify angle area in TEE slice images extracted from videos,the classification resuh was a priori medical probability;then,another classifier was trained to recognize TEE slice images and a conditional probability was obtained.Finally,the final recognition result was got by Bayesian approach.The experimental results indicate that the model based on deep learning and medical priori knowledge get higher accuracy than the traditional methods when recognizing TEE images.