Generative adversarial networks (GANs) has proven hugely successful, but suffer from train instability. The recently proposed Wasserstein GAN (WGAN) has largely overcome the problem, but can still fail to converge in some case or be to complex. It has been found that the use of weight clipping in WGAN to enforce a Lipschitz constraint on the critic, is the cause of the failure. We modify network architecture: use domain constraint layer instead of the use of weight clipping in WGAN. Experimental results show that our proposed method generates higher quality images than WGAN with weight clipping. And architecture is sample. Beside the network is more stable and easier to train.
Segmentation of anatomical structures in ultrasound images required radiological technology and a great deal of ultrasonic experience. The manual segmentation is often dependent on expertise of clinicians and time-consuming. Therefore, we present an automatic system for segmentation and measurement of ultrasound images. We propose a scale attention feature pyramid network (SAFNet) for fetal biometric measurements from two-dimensional ultrasound images. The scale attention module is steered to form feature pyramid at each level. Auxiliary layer is used to learn object boundary definition with deep supervision. Further, we present a two-stage framework which is an automatic classification measurement system (ACMS), firstly classifies the image type which has three labels: head, abdomen and femur. Then outputs the final segmentation result. The SAFNet results better performance on our datasets compared to the baseline U-Net. Experiments show that the ACMS results in classification accuracy of 95.27%/90.94%/94.93% of fetal head, abdomen and femur test set, respectively. Feature pyramid and attention mechanism inside the network for feature selection results in improvement in the segmentation accuracy. The ACMS can conveniently obtain segmentation result no matter what type is given.
Generative adversarial networks (GANs) has achieved success in many fields. However, there are some samples generated by many GAN-based works, whose structure is ambiguous. In this work, we propose Structure Guided GANs that introduce structural similar into GANs to overcome the problem. In order to achieve our goal, we introduce an encoder and a decoder into a generator to design a new generator and take real samples as part of the input of a generator. And we modify the loss function of the generator accordingly. By comparison with WGAN, experimental results show that our proposed method overcomes largely sample structure ambiguous and can generate higher quality samples.
Image recognition technology has been widely applied and played an important role in various fields nowadays. Because of multi-layer structure of deep network can use a more concise way to express complex functions, deep neural network (DNN) will be applied to the image recognition to improve the accuracy of image classification. Analysis the existing problems of deep neural network. Then put forward new approaches to solve the gradient vanishing and over-fitting problems. The experimental results which verified on the MNIST, show that our proposed approaches can improve the classification accuracy greatly and accelerate the convergence speed. Compared to support vector machine (SVM), the optimized model of the neural network is not only effective, but also converged quickly.