The X-ray radiation from computed tomography (CT) brought us the potential risk. Simply decreasing the dose makes the CT images noisy and diagnostic performance compromised. Here, we develop a novel denoising low-dose CT image method. Our framework is based on an improved generative adversarial network coupling with the hybrid loss function, including the adversarial loss, perceptual loss, sharpness loss, and structural similarity loss. Among the loss function terms, perceptual loss and structural similarity loss are made use of to preserve textural details, and sharpness loss can make reconstruction images clear. The adversarial loss can sharp the boundary regions. The results of experiments show the proposed method can effectively remove noise and artifacts better than the state-of-the-art methods in the aspects of the visual effect, the quantitative measurements, and the texture details.

A novel reconstruction algorithm is presented to address the noise artifacts of path tracing. SURE (Stein’s unbiased risk estimator) is adopted to estimate the noise level per pixel that guides adaptive sampling process. Modified MLPs (multilayer perceptron) network is used to predict the optimal reconstruction parameters. In sampling stage, coarse samples are firstly generated. Then each noise level is estimated with SURE. Additional samples are distributed to the pixels with high noise level. Next, we extract a few features from the results of adaptive sampling used for the subsequent reconstruction stage. In reconstruction stage, modified MLPs network is adopted to model a complex relationship between extracted features and optimal reconstruction parameters. An anisotropic filter is used to reconstruct the final images with the parameters predicted by neural networks. Compared to the state-of-the-art methods, experiment results demonstrate that our algorithm performs better than other methods in numerical error and visual image quality.

Brain tumors are characterized by their relatively high incidence and mortality rates, highlighting the utmost importance of precise automatic segmentation for subsequent diagnosis and treatment. Although deep learning has significantly advanced the field of accurate and efficient automatic brain tumor segmentation, several challenges still persist. In this paper, we introduce a novel architecture called the Dual Encoder Mirror Difference Residual U-Net (DEMD-ResUNet). This approach incorporates dual encoders that process both the original and horizontally flipped images. Additionally, residual blocks are employed to substitute the original convolutional blocks in the encoder section of the U-Net structure. This modification not only streamlines network training but also mitigates issues related to network degradation and the loss of detailed information. To further enhance feature representation, we propose a Multimodal Difference Feature Augmentation (MDFA) module, which effectively highlights abnormal regions in both the original and mirrored brain tumor images to facilitate better feature discrimination. Moreover, a Mirror Difference Feature Fusion (MDFF) module is integrated between the dual encoders and the decoder. This module efficiently transfers features from both the original and mirrored images to the decoder, leveraging the symmetrical information inherent in the images and subsequently boosting the segmentation performance of the model. Ablation experiments conducted on the DEMD-ResUNet model demonstrate the efficacy of its various modules and hyperparameter settings. When evaluated on the BraTS 2018 and BraTS 2019 datasets, our model achieves impressive Dice similarity coefficient (DSC) values of 0.862, 0.925, and 0.905 for Enhanced tumor (ET), Whole tumor (WT), and Tumor core (TC) in the former, and 0.869, 0.922, and 0.916 in the latter, respectively.

Post-stroke neuroimaging is the key to the treatment of brain stroke. Typically, segmenting lesions manually is not only time-consuming, but limited by the varying morphology of lesions and the similarity of tissue intensity distribution. In recent years, with the rapidly and widely applying of deep learning technology in medical imaging, it becomes the one of the hot-spots. However, most stroke segmentation techniques could not utilize the structural symmetry information efficiently. Moreover, there are several problems, such as diverse lesions location, large changes of scale and unclear boundaries of lesions. To address these shortcomings, this paper presents a novel stroke segmentation model, called BSSNet. Firstly, the standard convolution of U-Net is replaced by the Depth-Wise Separable Convolution(DWSC) and the residual connection operation is used to reduce the loss of feature information in the encoder and decoder. Secondly, between the encoder and decoder, this paper introduces the Multi-Scale Channel Attention(MSCA) module to effectively improve the segmentation performance. Finally, the Attention-Guided Connection(AGC) module takes place of the original connection operation, which can select more context information from features with low-level guided by the features with high-level. To demonstrate the advantages of the proposed BSSNet, the comparison experiments are conducted on the open Anatomical Tracings of Lesions After Stroke(ATLAS) datasets. Experiment results shows that our model outperforms the state-of-the-art segmentation methods both in the quantitative metrics and visual effects.

In the global outbreak of corona-virus disease (COVID-19), it is of foremost priority to find an efficient and faster diagnosis method to reduce the transmission rate of the COVID-19 disease. Recently, deep learning-based detection methods have been proposed as a potential diagnostic tool, because there is much essential information in the radio-logical images for COVID-19 detection. This paper intends to develop a hybrid deep neural network (COVID19-ResCapsNet), using the X-ray images, to predict the risk at the onset of disease in patients suffering from COVID-19. The proposed COVID19-ResCapsNet method consists of two steps. The first step is mainly to extract features with the improved Residual feature extraction network. In the final step, several Capsule Networks are employed to classify COVID-19, and Non COVID-19. The proposed method is evaluated on two CXR image datasets(Dataset-1: 4236 images and Dataset-2: 2250 images) and the classification accuracy can reach 0.9988 on Dataset-1 and 0.9933 on Dataset-2 for binary classification cases. The results are compared with several state-of-the-art pre-trained classification models. The effects of two important hyper parameters: different optimizer and batch size, have also been investigated. Hence, the proposed COVID-19 detection model is effective and can be used as the supplementary tool in clinical application.