Abstract
Automatic segmentation of lumen contour plays an important role in medical imaging and diagnosis, which is the first step towards the evaluation of morphology of vessels under analysis and the identification of possible atherosclerotic lesions. Meanwhile, quantitative information can only be obtained with segmentation, contributing to the appearance of novel methods which can be successfully applied to intravascular optical coherence tomography (IVOCT) images. This paper proposed a new end-to-end neural network (N-Net) for the automatic lumen segmentation, using multi-scale features based deep neural network, for IVOCT images. The architecture of the N-Net contains a multi-scale input layer, a N-type convolution network layer and a cross-entropy loss function. The multi-scale input layer in the proposed N-Net is designed to avoid the loss of information caused by pooling in traditional U-Net and also enriches the detailed information in each layer. The N-type convolutional network is proposed as the framework in the whole deep architecture. Finally, the loss function guarantees the degree of fidelity between the output of proposed method and the manually labeled output. In order to enlarge the training set, data augmentation is also introduced. We evaluated our method against loss, accuracy, recall, dice similarity coefficient, jaccard similarity coefficient and specificity. The experimental results presented in this paper demonstrate the superior performance of the proposed N-Net architecture, comparing to some existing networks, for enhancing the precision of automatic lumen segmentation and increasing the detailed information of edges of the vascular lumen.
Highlights
Coronary heart disease has become a widespread health concern causing fatality worldwide, known as a myocardial ischemic heart disease caused by coronary artery stenosis or obstruction
In view of the problems encountered in the above segmentation methods, we propose an automatic segmentation algorithm for vascular cavity contours using deep neural networks based on multi-scale input fusion
The N-Net consists of three parts: the first part is multi-scale input layer, which is used to construct image pyramid input to achieve multi-level receptive field fusion; the second part is N-type convolution neural network, which can extract multi-level image features; the third part is the cross-entropy loss function, which is used to compare the difference between the experimental and manual results
Summary
Coronary heart disease has become a widespread health concern causing fatality worldwide, known as a myocardial ischemic heart disease caused by coronary artery stenosis or obstruction. The U-Net network is contained as the main body of the proposed N-Net. The N-Net consists of three parts: the first part is multi-scale input layer, which is used to construct image pyramid input to achieve multi-level receptive field fusion; the second part is N-type convolution neural network, which can extract multi-level image features; the third part is the cross-entropy loss function, which is used to compare the difference between the experimental and manual results. If the new lumen images with different resolution are fused directly with the feature graph of N-Net fusion, the network parameters will be too much For this purpose, the fusion is divided into two steps, and the dimension of fused feature graph is reduced to 1 by 1 convolution, contributing to reduction of the parameters and increment of the information interaction between features. In order to get the probability of each pixel belonging to the category, Sigmoid function is used in the last layer for activation, ensuring that the output probability value is between 0 and 1
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call