基于深度神经网络的冠脉造影图像的血管狭窄自动定位及分类预测

提出一套基于深度神经网络与监督学习的算法,用于对冠状动脉图像中的血管狭窄特征进行自动检测和分类.主要利用冠脉造影定量分析(QCA)作为标签进行监督学习,将冠脉狭窄的严重程度分为正常(＜25％狭窄分数)、狭窄(＞25％狭窄)类别,并实现图像中的狭窄定位检测.利用inception模型作为基础分类器,对图像级狭窄进行初步分...

Automatic Location and Classification of Coronary Artery Stenosis Based on Deep Neural Network

In this paper, a deep neural network-based workflow was proposed to automatically detect and classify the stenosis features in coronary images. The algorithm mainly used quantitative coronary angiography (QCA) as a label for supervised learning and classifies the severity of coronary stenosis into normal (< 25% stenosis score) and stenosis (> 25% stenosis) categories and realized stenosis location detection in images. The algorithm used the inception model as the basic classifier to preliminarily classify the image level stenosis, and then combined with the multi-level pool structure to jointly predict the multi perspective angiography image to obtain the left-/right-artery/patient level stenosis prediction. On the basis of the classifier, the feature was further extracted, and the unsupervised learning model was used to realize the narrow location in the image.The training and cross validation were performed on a total of 10872 images in 235 clinical studies. The results showed that the algorithm achieved 85% accuracy and 0.91 AUC score in image level stenosis classification; in multi view joint prediction experiment, the sensitivity and AUC score of 0.94/0.90/0.96 and 0.87/0.88/0.86 respectively for left-/ right-/patient level stenosis classification prediction. In the stenosis localization experiment, the sensitivity of detection for left-/right-artery stenosis was 0.70/0.68, and the mean square error of 512 × 512 image was 37.6/39.3 pixels, respectively. In conclusion, the proposed method realized the potential of auxiliary diagnosis prediction from image to patient with high accuracy, which not only provided the preliminary screening ability in the process of coronary angiography, but also laid the foundation for more accurate and automatic computer-aided diagnosis.