Heart segmentation is challenging due to the poor image contrast of heart in the CT images. Since manual segmentation of the heart is tedious and time-consuming, we propose an attention-based Convolution Neural Network (CNN) for heart segmentation. First, one-hot preprocessing is performed on the multi-tissue CT images. U-Net network with Attention-gate is then applied to obtain the heart region. We compared our method with several CNN methods in terms of dice coefficient. Results show that our method outperforms other methods for segmentation.
Abstract Diabetes interferes with the body’s ability to use and store sugar (glucose), which it can cause damage throughout the body. Generally, Microaneurysm (MA) is the first clinically observable lesions in the Diabetic Retinopathy (DR). Since interference of biological objects in eye fundus images are similar to MA, locate Microaneurysm lesions precisely is a challenging task for researches. This research employs U-Net Convolutional Neural Network to model saliency of objects in images, global and local context are both taken into account to provide a better initialization for the process. Feature extraction techniques are applied first to assign a local saliency value to each pixel by considering its local context from fundus images such as ORB, SURF, and MSER. The extracted feature vectors are applied for training the network. The sum of the weighted salient object regions produce the final saliency map, then implements U-Net to segment MA lesions.Our experiment has carried out using the publicly available Indian Diabetic Retinopathy Image Dataset (IDRiD), which has used in "Diabetic Retinopathy: Segmentation and Grading Challenge" workshop and our proposed method has given an outstanding accuracy of 98.78 %.
Abstract Pulmonary vein anatomical structure typing plays a crucial role in the preoperative assessment and postoperative evaluation of lung tumor resection, atrial fibrillation radio frequency ablation, and other medical procedures. The accuracy of such typing relies heavily on the segmentation results of the left atrium and proximal pulmonary veins. However, due to the similarities in intensity between the left atrium, proximal pulmonary veins, and adjacent tissues in CT images, segmentation errors often occur, leading to subsequent inaccuracies in pulmonary vein classification. To address this issue, we propose an attention module called Dimensional Decomposition Attention (DDA), which combines Dimensional Decomposition Spatial Attention (DDSA) and Dimensional Decomposition Channel Attention (DDCA). DDA effectively leverages the spatial and channel information of 3D images to enhance the segmentation accuracy of the left atrium and proximal pulmonary veins. In DDSA, the input features are decomposed into three one‐dimensional directional features (height, width, and depth) and fused to generate weights that emphasize spatial shape features and focus on the region of interest. On the other hand, DDCA encodes the input features into dimensional channel features, fuses them with one‐dimensional directional features, and utilizes position encoding to reinforce the channel features and prioritize channels with relevant information. The performance of DDA was evaluated using a two‐stage experimental approach on datasets provided by The People's Hospital of Liaoning Province and the MM‐WHS CT dataset, yielding average Dice values of 93.93% and 90.80%, respectively, demonstrating the effectiveness of DDA.
Abstract Breast cancer is the most common cancer for women, and it also is the leading cause of cancer‐related deaths. As a highly heterogeneous disease, it is crucial to identify the molecular subtypes of breast cancer before individualized therapy. Therefore, we proposed a multiloss network framework to classify the breast cancer subtype based on digital breast tomosynthesis (DBT) images. We employed the multiloss strategy to learn the low‐level features more efficiently and effectively, which would provide a better basis for extracting the high‐level features. Additionally, we also proposed a decomposed attention block (DA), which not only captured the interdependencies between all channels but also the precise positional information at the x‐ and y‐dimensions. The multiloss strategy enables the network to learn useful feature representations of breast cancer subtypes, and the DA block further improves the classification performance by capturing the information between channels and positions.
It is a significant problem to learn the scene correlation of uncalibrated static cameras, which can be applied for intelligent surveillance systems with a large scale camera network. Some existing approaches learn the scene correlation among camera views by tracking targets across cameras. They seldom analyze the scene correlation among cameras with appearance modeling for moving targets. In this paper, a novel adaptive appearance co-occurrence modeling approach is proposed to learn the scene correlation by long term statistics. Firstly, spatial-depth scale-invariant model (SDSI) is introduced to make a scale normalization for the whole camera views. A uniform metric of target scales is established in the system so that an absolute height is obtain to make an identification of moving targets with similar appearance. Then, the appearance co-occurrence modeling is formulated to learn the spatio-temporal co-occurrence relationship among cameras with detection for targets with same appearance in temporal neighbourhood. The proposed approach generates visual attention cross a number of camera views in case that the cameras are not calibrated, which is adaptive to learn the co-occurrence correlation. The effectiveness of our approach is demonstrated in PKU-SES system with 10 cameras in two sites.
<abstract> <p>Accurate abdomen tissues segmentation is one of the crucial tasks in radiation therapy planning of related diseases. However, abdomen tissues segmentation (liver, kidney) is difficult because the low contrast between abdomen tissues and their surrounding organs. In this paper, an attention-based deep learning method for automated abdomen tissues segmentation is proposed. In our method, image cropping is first applied to the original images. U-net model with attention mechanism is then constructed to obtain the initial abdomen tissues. Finally, level set evolution which consists of three energy terms is used for optimize the initial abdomen segmentation. The proposed model is evaluated across 470 subsets. For liver segmentation, the mean dice are 96.2 and 95.1% for the FLARE21 datasets and the LiTS datasets, respectively. For kidney segmentation, the mean dice are 96.6 and 95.7% for the FLARE21 datasets and the LiTS datasets, respectively. Experimental evaluation exhibits that the proposed method can obtain better segmentation results than other methods.</p> </abstract>
The existing CNN-based segmentation methods use the object regions alone as the labels to train their networks, and the potentially useful boundaries annotated by radiologists are not used directly during the training.Thus, we proposed a framework of double U-Nets to integrate object regions and boundaries for more accurate segmentation.The proposed network consisted of a down-sampling path followed by two symmetric up-sampling paths.The down-sampling path learned the low-level features of regions and boundaries, and two up-sampling paths learned the high-level features of regions and boundaries, respectively.The outputs from the down-sampling path were concatenated with the corresponding ones from two up-sampling paths by skip connections.The outputs of double U-Nets were the predicted probability images of object regions and boundaries, and they were integrated to calculate the dice loss with the corresponding labels.The proposed double U-Nets were evaluated on two datasets: 247 radiographs for the segmentation of lungs, hearts, and clavicles, and 284 radiographs for the segmentation of pelvises.Compared with the baseline U-Net, our double U-Nets improved the mean dices and reduced the 90% Hausdorff distances for the ''difficult'' objects (lower lungs, clavicles, and pelvises), and the integration of ''difficult'' object regions and boundaries can improve the segmentation results compared with the use of object regions alone.However, for the ''easy'' objects (entire lungs and hearts) or ''very difficult'' objects (pelvises in lateral and implanted images), the integration did not improve the segmentation performance.
This paper presents a new real-time localization method for indoor service robots based on camera and laser. First of all, it is proved that planar motion of an upward camera in the floor plane can be solved by only two matching points with equal depth. Then, a real-time indoor localization method is proposed accordingly using an upward fisheye camera mounted on the robot and two fixed laser pointers which project red and green laser spots onto the ceiling. In addition, to improve the localization accuracy, an Extended Kalman Filter is employed with active matching to speed up laser spots recognition process. Finally, the proposed method is implemented on a mobile developer board, and tested with P3-DX platform in indoor environment. Experimental results show the excellent performance of the given method.
'/%3e%3cuse xlink:href='%23a' transform='rotate(23.036 .093 25.536)'/%3e%3cuse xlink:href='%23a' transform='rotate(45.87 1.273 16.18)'/%3e%3cuse xlink:href='%23a' transform='rotate(69.945 .996 12.078)'/%3e%3cuse xlink:href='%23a' transform='rotate(20.66 -19.689 31.932)'/%3e%3c/svg%3e)