The deep learning methods for various disease prediction tasks have become very effective and even surpass human experts. However, the lack of interpretability and medical expertise limits its clinical application. Th...
Purpose: We aim to develop a back-propagation artificial neural network (BP-ANN) improved by a priori knowledge and to compare its efficacy with other methods in early diabetic retinopathy (DR) detection.Methods: A to...
Purpose: We aim to develop a back-propagation artificial neural network (BP-ANN) improved by a priori knowledge and to compare its efficacy with other methods in early diabetic retinopathy (DR) detection.
Methods: A total of 240 fundus images, composed of 120 early-stage DR and 120 normal images, were obtained with the same 45° field of view camera, with the macula at the center, as a cohort for further training. All retinal images were processed, and a priori knowledge features such as blood vessel width and tortuosity were semi-automatically extracted. An improved BP-ANN with a priori knowledge was developed, and its efficacy was compared with that of the traditional BP network and SVM. Besides, k-fold cross validation method was conducted to demonstrate the efficiency of the proposed methods. We also developed a graphical user interface of our proposed BP-ANN to aid in DR screening.
Results: Our 10 randomization and 5-fold cross validation results of SVM, traditional BP, and improved BP were compared. The results indicated that the BP-ANN with a priori knowledge can achieve better detection results. Besides, our results were also comparable with other reported state-of-art algorithms. During the training stage, the epoch in the improved BP-ANN was less than that in the traditional BP group (109 vs 254), indicating that the time cost was shorter when using our improved BP-ANN. Furthermore, the accuracy and epoch of both the traditional BP and our improved BP network obtained better performances when the number of hidden neurons was 20.
Conclusions: A priori knowledge-based BP-ANN could be a promising measure for early DR detection.
CCS: Information system→Expert system
According to fractional order model, the governing equation of viscoelastic plate is established. Using the properties of the shifted Bernstein polynomials, the fractional order equation of viscoelastic plate is resol...
详细信息
Complex networks have attracted growing attention in many fields. As a generalization of fractal analysis, multifractal analysis (MFA) is a useful way to systematically describe the spatial heterogeneity of both theor...
Complex networks have attracted growing attention in many fields. As a generalization of fractal analysis, multifractal analysis (MFA) is a useful way to systematically describe the spatial heterogeneity of both theoretical and experimental fractal patterns. Some algorithms for MFA of unweighted complex networks have been proposed in the past a few years, including the sandbox (SB) algorithm recently employed by our group. In this paper, a modified SB algorithm (we call it SBw algorithm) is proposed for MFA of weighted networks. First, we use the SBw algorithm to study the multifractal property of two families of weighted fractal networks (WFNs): "Sierpinski" WFNs and "Cantor dust" WFNs. We also discuss how the fractal dimension and generalized fractal dimensions change with the edge-weights of the WFN. From the comparison between the theoretical and numerical fractal dimensions of these networks, we can find that the proposed SBw algorithm is efficient and feasible for MFA of weighted networks. Then, we apply the SBw algorithm to study multifractal properties of some real weighted networks - collaboration networks. It is found that the multifractality exists in these weighted networks, and is affected by their edge-weights.
As people come into contact with image data more often, high quality and clear images attract more attention. Many methods have been proposed to deal with image noise problem including deep learning (DL). However most...
As people come into contact with image data more often, high quality and clear images attract more attention. Many methods have been proposed to deal with image noise problem including deep learning (DL). However most of them is lack of capability when customers want more perceptual details of the image without information loss. In this paper, a deep residual network based on generative adversarial (GAN) network was proposed to complete the image denoising mission. Firstly, a generative-adversarial network structure based on residual blocks was designed. Secondly, a refined loss function was given to train the GAN network. The well designed loss function can help the generated image to be very close to the clear counterpart (ground truth) while enhancing more details in colours and brightness. Finally, extensive experiments show that our network is not only convincing for images denoising, but also effective for other image process tasks, such as image defogging, medical CT denoising etc., presenting impressive and competitive effects.
Deep reinforcement learning (deep RL) achieved big successes with the advantage of deep learning techniques, while it also introduces the disadvantage of the model interpretability. Bad interpretability is a great obs...
Deep reinforcement learning (deep RL) achieved big successes with the advantage of deep learning techniques, while it also introduces the disadvantage of the model interpretability. Bad interpretability is a great obstacle for deep RL to be applied in real situations or human-machine interaction situations. Borrowed from the deep learning field, the techniques of saliency maps recently become popular to improve the interpretability of deep RL. However, the saliency maps still cannot provide specific and clear enough model interpretations for the behavior of deep RL agents. In this paper, we propose to use hierarchical conceptual embedding techniques to introduce prior-knowledge in the deep neural network (DNN) based models of deep RL agents and then generate the saliency maps for all the embedded factors. As a result, we can track and discover the important factors that influence the decisions of deep RL agents.
Through MATLAB, the paper makes a comparison between Principal Component Analysis (PCA)face recognition algorithm and Adaboost recognition algorithm and selects the algorithm with higher recognition rates to develop a...
Through MATLAB, the paper makes a comparison between Principal Component Analysis (PCA)face recognition algorithm and Adaboost recognition algorithm and selects the algorithm with higher recognition rates to develop an auto face recognition system. The paper explicates primary techniques the system adopts and its specific realization process. By downloading face database online, the paper conducts an all-round test to the system, the result of which proves that this face recognition system is completely practical and feasible.
The Area Under the ROC Curve (AUC) is a crucial metric for machine learning, which evaluates the average performance over all possible True Positive Rates (TPRs) and False Positive Rates (FPRs). Based on the knowledge...
详细信息
暂无评论