Automatic and accurate cervical nucleus segmentation is important because nuclei carry substantial diagnostic information for automatic computer-assisted cervical cancer screening and diagnosis systems. In this paper, we propose a cervical nucleus segmentation method in which pixel-level prior information is utilized to provide the supervisory information for the training of a mask regional convolutional neural network (Mask-RCNN), which is then employed to extract the multi-scale features of the nuclei, and the coarse segmentation and bounding box of the nuclei are obtained by forward propagation of the Mask-RCNN. To refine the segmentation, a local fully connected conditional random field (LFCCRF) that contains unary and pairwise energy terms is employed. The nuclear region of interest is determined by extending the bounding box, the coarse segmentation in the nuclear region is used to construct the unary energy, and the pairwise energy is contributed by the position and intensity information of all of the pixels in the nuclear region. By minimizing the energy of the LFCCRF, the final segmentation is realized. We evaluated our method by using cervical nuclei from the Herlev Pap smear data set in this paper, and the precision, recall, and Zijdenbos similarity index were all found to be greater than 0.95 with low standard deviations, demonstrating that our method enables more accurate and stable cervical nucleus segmentation than the current state-of-the-art methods.

Generative adversarial networks (GAN) have shown great potential for image quality improvement in low-dose CT (LDCT). In general, the shallow features of generator include more shallow visual information such as edges and texture, while the deep features of generator contain more deep semantic information such as organization structure. To improve the network's ability to categorically deal with different kinds of information, this paper proposes a new type of GAN with dual-encoder- single-decoder structure. In the structure of the generator, firstly, a pyramid non-local attention module in the main encoder channel is designed to improve the feature extraction effectiveness by enhancing the features with self-similarity; Secondly, another encoder with shallow feature processing module and deep feature processing module is proposed to improve the encoding capabilities of the generator; Finally, the final denoised CT image is generated by fusing main encoder's features, shallow visual features, and deep semantic features. The quality of the generated images is improved due to the use of feature complementation in the generator. In order to improve the adversarial training ability of discriminator, a hierarchical-split ResNet structure is proposed, which improves the feature's richness and reduces the feature's redundancy in discriminator. The experimental results show that compared with the traditional single-encoder- single-decoder based GAN, the proposed method performs better in both image quality and medical diagnostic acceptability. Code is available in https://github.com/hanzefang/DESDGAN .

Convolutional neural network (CNN)-based deep learning techniques have enjoyed many successful applications in the field of medical imaging. However, the complicated between-manifold projection from the projection domain to the spatial domain hinders the direct application of CNN techniques in computed tomography (CT) reconstruction. In this work, we proposed a novel CT reconstruction framework based on a CNN, i.e., an intelligent back-projection network (iBP-Net). The proposed iBP-Net method fused a pre-CNN, a back-projection layer, and a post-CNN into an end-to-end network. The pre-CNN adopted CNN techniques to model a filtering operation in the projection domain. In the back-projection layer, a back-projection operation was employed to perform between-manifold projection. Based on CNN techniques, the post-CNN worked together with the pre-CNN to recover reconstructed images from the outputs of the back-projection layer in the spatial domain while maintaining high visual sensitivity. The experimental results demonstrate the feasibility of the proposed iBP-Net framework for CT reconstruction.

The 3D sparse-view computed laminography (CL) is an important research issue due to its considerable potential in lowering costs. To solve the artifacts and noise caused by sparse-view scanning, the total variation (TV) method is widely used. However, TV only utilizes the prior information of the reconstructed image itself. Therefore, the Prior-image-based reconstruction method (PIBR) is proposed to reduce the artifacts by using a prior image. In this study, we proposed a 3D CL method based on the final prior image and TV for sparse-view scanning, which is divided into two sub-problems corresponding to two solution steps: the novel PIBR step and the denoising step. In the first step, we devise the final prior image and its corresponding prior mask to ensure the effectiveness of the PIBR. The final prior image is obtained based on the prior images, which are reconstructed from two different oblique angle projections. The prior mask is a binary mask, which is used to record the position of the prior image information. Then the denoising step is based on the TV minimization method using the Split-Bregman (SB) frame. Soft shrinkage operations and Fast Fourier Transform (FFT) are used to efficiently implement the SB frame. Compared with several iterative reconstruction methods, the experimental results demonstrate the effectiveness of the proposed method in terms of preserving edges, suppressing inter-slices aliasing and denoising.

Direct aperture optimization (DAO) is an effective method to generate high-quality intensitymodulated radiation therapy treatment plans. In generic DAO, the direction of negative gradient descent is generally used to determine the aperture shape. However, this strategy can reduce the convergence rate, especially near the optimal value. We propose aperture shape generation based on the direction of gradient descent with momentum, where column generation is implemented as carrier. During aperture shape generation of column generation, the current aperture gradient map is first calculated. Then, the gradient with momentum is calculated based on the existing gradient information. Finally, the direction of gradient descent with momentum is constructed for obtaining the deliverable aperture shape by solving the pricing problem. To verify the effectiveness of the proposed method, we conducted comparative experiments on two head and neck and two prostate tumor cases. Compared with generic column generation, the proposed method can effectively protect the organs at risk while ensuring the required dose distribution to the target. Using the proposed method, the number of apertures and optimization time can be reduced by up to 30.95 and 32.96%, respectively, compared to the conventional approach. The experimental results suggest that the proposed method can accelerate the search speed and improve the quality of treatment plans.