Given the potential risk of X-ray radiation to the patient, low-dose CT has attracted a considerable interest in the medical imaging field. Currently, the main stream low-dose CT methods include vendor-specific sinogram domain filtration and iterative reconstruction algorithms, but they need to access raw data, whose formats are not transparent to most users. Due to the difficulty of modeling the statistical characteristics in the image domain, the existing methods for directly processing reconstructed images cannot eliminate image noise very well while keeping structural details. Inspired by the idea of deep learning, here we combine the autoencoder, deconvolution network, and shortcut connections into the residual encoder-decoder convolutional neural network (RED-CNN) for low-dose CT imaging. After patch-based training, the proposed RED-CNN achieves a competitive performance relative to the-state-of-art methods in both simulated and clinical cases. Especially, our method has been favorably evaluated in terms of noise suppression, structural preservation, and lesion detection.

Photon-counting detectors can acquire x-ray intensity data in different energy bins. The signal-to-noise ratio of resultant raw data in each energy bin is generally low due to the narrow bin width and quantum noise. To address this problem, here we propose an image reconstruction approach for spectral CT to simultaneously reconstruct x-ray attenuation coefficients in all the energy bins. Because the measured spectral data are highly correlated among the x-ray energy bins, the intraimage sparsity and interimage similarity are important prior knowledge for image reconstruction. Inspired by this observation, the total variation and spectral mean measures are combined to improve the quality of reconstructed images. For this purpose, a linear mapping function is used to minimalize image differences between energy bins. The split Bregman technique is applied to perform image reconstruction. Our numerical and experimental results show that the proposed algorithms outperform competing iterative algorithms in this context.

The main task of magnetic resonance imaging (MRI) automatic brain tumor segmentation is to automatically segment the brain tumor edema, peritumoral edema, endoscopic core, enhancing tumor core and nonenhancing tumor core from 3D MR images. Because the location, size, shape and intensity of brain tumors vary greatly, it is very difficult to segment these brain tumor regions automatically. In this paper, by combining the advantages of DenseNet and ResNet, we proposed a new 3D U-Net with dense encoder blocks and residual decoder blocks. We used dense blocks in the encoder part and residual blocks in the decoder part. The number of output feature maps increases with the network layers in contracting path of encoder, which is consistent with the characteristics of dense blocks. Using dense blocks can decrease the number of network parameters, deepen network layers, strengthen feature propagation, alleviate vanishing-gradient and enlarge receptive fields. The residual blocks were used in the decoder to replace the convolution neural block of original U-Net, which made the network performance better. Our proposed approach was trained and validated on the BraTS2019 training and validation data set. We obtained dice scores of 0.901, 0.815 and 0.766 for whole tumor, tumor core and enhancing tumor core respectively on the BraTS2019 validation data set. Our method has the better performance than the original 3D U-Net. The results of our experiment demonstrate that compared with some state-of-the-art methods, our approach is a competitive automatic brain tumor segmentation method.

Activity recognition is a fundamental and crucial task in computer vision. Impressive results have been achieved for activity recognition in high-resolution videos, but for extreme low-resolution videos, which capture the action information at a distance and are vital for preserving privacy, the performance of activity recognition algorithms is far from satisfactory. The reason is that extreme low-resolution (e.g., 12 × 16 pixels) images lack adequate scene and appearance information, which is needed for efficient recognition. To address this problem, we propose a super-resolution-driven generative adversarial network for activity recognition. To fully take advantage of the latent information in low-resolution images, a powerful network module is employed to super-resolve the extremely low-resolution images with a large scale factor. Then, a general activity recognition network is applied to analyze the super-resolved video clips. Extensive experiments on two public benchmarks were conducted to evaluate the effectiveness of our proposed method. The results demonstrate that our method outperforms several state-of-the-art low-resolution activity recognition approaches.