A dense and U-shaped transformer with dual-domain multi-loss function for sparse-view CT reconstruction

OBJECTIVE

CT image reconstruction from sparse-view projections is an important imaging configuration for low-dose CT, as it can reduce radiation dose. However, the CT images reconstructed from sparse-view projections by traditional analytic algorithms suffer from severe sparse artifacts. Therefore, it is of great value to develop advanced methods to suppress these artifacts. In this work, we aim to use a deep learning (DL)-based method to suppress sparse artifacts.

METHODS

Inspired by the good performance of DenseNet and Transformer architecture in computer vision tasks, we propose a Dense U-shaped Transformer (D-U-Transformer) to suppress sparse artifacts. This architecture exploits the advantages of densely connected convolutions in capturing local context and Transformer in modelling long-range dependencies, and applies channel attention to fusion features. Moreover, we design a dual-domain multi-loss function with learned weights for the optimization of the model to further improve image quality.

RESULTS

Experimental results of our proposed D-U-Transformer yield performance improvements on the well-known Mayo Clinic LDCT dataset over several representative DL-based models in terms of artifact suppression and image feature preservation. Extensive internal ablation experiments demonstrate the effectiveness of the components in the proposed model for sparse-view computed tomography (SVCT) reconstruction.

SIGNIFICANCE

The proposed method can effectively suppress sparse artifacts and achieve high-precision SVCT reconstruction, thus promoting clinical CT scanning towards low-dose radiation and high-quality imaging. The findings of this work can be applied to denoising and artifact removal tasks in CT and other medical images.

A new architecture combining convolutional and transformer-based networks for automatic 3D multi-organ segmentation on CT images

PURPOSE

Deep learning-based networks have become increasingly popular in the field of medical image segmentation. The purpose of this research was to develop and optimize a new architecture for automatic segmentation of the prostate gland and normal organs in the pelvic, thoracic, and upper gastro-intestinal (GI) regions.

METHODS

We developed an architecture which combines a shifted-window (Swin) transformer with a convolutional U-Net. The network includes a parallel encoder, a cross-fusion block, and a CNN-based decoder to extract local and global information and merge related features on the same scale. A skip connection is applied between the cross-fusion block and decoder to integrate low-level semantic features. Attention gates (AGs) are integrated within the CNN to suppress features in image background regions. Our network is termed "SwinAttUNet." We optimized the architecture for automatic image segmentation. Training datasets consisted of planning-CT datasets from 300 prostate cancer patients from an institutional database and 100 CT datasets from a publicly available dataset (CT-ORG). Images were linearly interpolated and resampled to a spatial resolution of (1.0 × 1.0× 1.5) mm3 . A volume patch (192 × 192 × 96) was used for training and inference, and the dataset was split into training (75%), validation (10%), and test (15%) cohorts. Data augmentation transforms were applied consisting of random flip, rotation, and intensity scaling. The loss function comprised Dice and cross-entropy equally weighted and summed. We evaluated Dice coefficients (DSC), 95th percentile Hausdorff Distances (HD95), and Average Surface Distances (ASD) between results of our network and ground truth data.

RESULTS

SwinAttUNet, DSC values were 86.54 ± 1.21, 94.15 ± 1.17, and 87.15 ± 1.68% and HD95 values were 5.06 ± 1.42, 3.16 ± 0.93, and 5.54 ± 1.63 mm for the prostate, bladder, and rectum, respectively. Respective ASD values were 1.45 ± 0.57, 0.82 ± 0.12, and 1.42 ± 0.38 mm. For the lung, liver, kidneys and pelvic bones, respective DSC values were: 97.90 ± 0.80, 96.16 ± 0.76, 93.74 ± 2.25, and 89.31 ± 3.87%. Respective HD95 values were: 5.13 ± 4.11, 2.73 ± 1.19, 2.29 ± 1.47, and 5.31 ± 1.25 mm. Respective ASD values were: 1.88 ± 1.45, 1.78 ± 1.21, 0.71 ± 0.43, and 1.21 ± 1.11 mm. Our network outperformed several existing deep learning approaches using only attention-based convolutional or Transformer-based feature strategies, as detailed in the results section.

CONCLUSIONS

We have demonstrated that our new architecture combining Transformer- and convolution-based features is able to better learn the local and global context for automatic segmentation of multi-organ, CT-based anatomy.

A Local and Non-Local Features Based Feedback Network on Super-Resolution

Recent advances in Single Image Super-Resolution (SISR) achieved a powerful reconstruction performance. The CNN-based network (both sequential-based and feedback-based) performs well in local features, while the self-attention-based network performs well in non-local features. However, single block cannot always perform well due to the realistic images always with multiple kinds of features. In order to take full advantage of different blocks on different features. We have chosen three different blocks cooperating to extract different kinds of features. Addressing this problem, in this paper, we propose a new Local and non-local features-based feedback network for SR (LNFSR): (1) The traditional deep convolutional network block is used to extract the local non-feedbackable information directly and non-local non-feedbackable information (needs to cooperate with other blocks). (2) The dense skip-based feedback block is use to extract local feedbackable information. (3) The non-local self-attention block is used to extract non-local feedbackable information and the based LR feature information. We also introduced the feature up-fusion-delivery blocks to help the features be delivered to the right block at the end of each iteration. Experiments show our proposed LNFSR can extract different kinds of feature maps by different blocks and outperform other state-of-the-art algorithms.

Real-Time Abnormal Object Detection for Video Surveillance in Smart Cities

With the adaptation of video surveillance in many areas for object detection, monitoring abnormal behavior in several cameras requires constant human tracking for a single camera operative, which is a tedious task. In multiview cameras, accurately detecting different types of guns and knives and classifying them from other video surveillance objects in real-time scenarios is difficult. Most detecting cameras are resource-constrained devices with limited computational capacities. To mitigate this problem, we proposed a resource-constrained lightweight subclass detection method based on a convolutional neural network to classify, locate, and detect different types of guns and knives effectively and efficiently in a real-time environment. In this paper, the detection classifier is a multiclass subclass detection convolutional neural network used to classify object frames into different sub-classes such as abnormal and normal. The achieved mean average precision by the best state-of-the-art framework to detect either a handgun or a knife is 84.21% or 90.20% on a single camera view. After extensive experiments, the best precision obtained by the proposed method for detecting different types of guns and knives was 97.50% on the ImageNet dataset and IMFDB, 90.50% on the open-image dataset, 93% on the Olmos dataset, and 90.7% precision on the multiview cameras. This resource-constrained device has shown a satisfactory result, with a precision score of 85.5% for detection in a multiview camera.

Incorporating the Breast Imaging Reporting and Data System Lexicon with a Fully Convolutional Network for Malignancy Detection on Breast Ultrasound

In this study, we applied semantic segmentation using a fully convolutional deep learning network to identify characteristics of the Breast Imaging Reporting and Data System (BI-RADS) lexicon from breast ultrasound images to facilitate clinical malignancy tumor classification. Among 378 images (204 benign and 174 malignant images) from 189 patients (102 benign breast tumor patients and 87 malignant patients), we identified seven malignant characteristics related to the BI-RADS lexicon in breast ultrasound. The mean accuracy and mean IU of the semantic segmentation were 32.82% and 28.88, respectively. The weighted intersection over union was 85.35%, and the area under the curve was 89.47%, showing better performance than similar semantic segmentation networks, SegNet and U-Net, in the same dataset. Our results suggest that the utilization of a deep learning network in combination with the BI-RADS lexicon can be an important supplemental tool when using ultrasound to diagnose breast malignancy.

Machine Learning Approach for Biopsy-Based Identification of Eosinophilic Esophagitis Reveals Importance of Global features

Goal

Eosinophilic esophagitis (EoE) is an allergic inflammatory condition characterized by eosinophil accumulation in the esophageal mucosa. EoE diagnosis includes a manual assessment of eosinophil levels in mucosal biopsies-a time-consuming, laborious task that is difficult to standardize. One of the main challenges in automating this process, like many other biopsy-based diagnostics, is detecting features that are small relative to the size of the biopsy.

Results

In this work, we utilized hematoxylin- and eosin-stained slides from esophageal biopsies from patients with active EoE and control subjects to develop a platform based on a deep convolutional neural network (DCNN) that can classify esophageal biopsies with an accuracy of 85%, sensitivity of 82.5%, and specificity of 87%. Moreover, by combining several downscaling and cropping strategies, we show that some of the features contributing to the correct classification are global rather than specific, local features.

Conclusions

We report the ability of artificial intelligence to identify EoE using computer vision analysis of esophageal biopsy slides. Further, the DCNN features associated with EoE are based on not only local eosinophils but also global histologic changes. Our approach can be used for other conditions that rely on biopsy-based histologic diagnostics.

Lightweight and Resource-Constrained Learning Network for Face Recognition with Performance Optimization

Despite considerable progress in face recognition technology in recent years, deep learning (DL) and convolutional neural networks (CNN) have revealed commendable recognition effects with the advent of artificial intelligence and big data. FaceNet was presented in 2015 and is able to significantly improve the accuracy of face recognition, while also being powerfully built to counteract several common issues, such as occlusion, blur, illumination change, and different angles of head pose. However, not all hardware can sustain the heavy computing load in the execution of the FaceNet model. In applications in the security industry, lightweight and efficient face recognition are two key points for facilitating the deployment of DL and CNN models directly in field devices, due to their limited edge computing capability and low equipment cost. To this end, this paper provides a lightweight learning network improved from FaceNet, which is called FN13, to break through the hardware limitation of constrained computational resources. The proposed FN13 takes the advantage of center loss to reduce the variations of the between-class features and enlarge the difference of the within-class features, instead of the triplet loss by using FaceNet. The resulting model reduces the number of parameters and maintains a high degree of accuracy, only requiring few grayscale reference images per subject. The validity of FN13 is demonstrated by conducting experiments on the Labeled Faces in the Wild (LFW) dataset, as well as an analytical discussion regarding specific disguise problems.

Automatic Pixel-Level Crack Detection on Dam Surface Using Deep Convolutional Network

Crack detection on dam surfaces is an important task for safe inspection of hydropower stations. More and more object detection methods based on deep learning are being applied to crack detection. However, most of the methods can only achieve the classification and rough location of cracks. Pixel-level crack detection can provide more intuitive and accurate detection results for dam health assessment. To realize pixel-level crack detection, a method of crack detection on dam surface (CDDS) using deep convolution network is proposed. First, we use an unmanned aerial vehicle (UAV) to collect dam surface images along a predetermined trajectory. Second, raw images are cropped. Then crack regions are manually labelled on cropped images to create the crack dataset, and the architecture of CDDS network is designed. Finally, the CDDS network is trained, validated and tested using the crack dataset. To validate the performance of the CDDS network, the predicted results are compared with ResNet152-based, SegNet, UNet and fully convolutional network (FCN). In terms of crack segmentation, the recall, precision, F-measure and IoU are 80.45%, 80.31%, 79.16%, and 66.76%. The results on test dataset show that the CDDS network has better performance for crack detection of dam surfaces.

Exploring Spatial Influence of Remotely Sensed PM2.5 Concentration Using a Developed Deep Convolutional Neural Network Model

Currently, more and more remotely sensed data are being accumulated, and the spatial analysis methods for remotely sensed data, especially big data, are desiderating innovation. A deep convolutional network (CNN) model is proposed in this paper for exploiting the spatial influence feature in remotely sensed data. The method was applied in investigating the magnitude of the spatial influence of four factors—population, gross domestic product (GDP), terrain, land-use and land-cover (LULC)—on remotely sensed PM2.5 concentration over China. Satisfactory results were produced by the method. It demonstrates that the deep CNN model can be well applied in the field of spatial analysing remotely sensed big data. And the accuracy of the deep CNN is much higher than of geographically weighted regression (GWR) based on comparation. The results showed that population spatial density, GDP spatial density, terrain, and LULC could together determine the spatial distribution of PM2.5 annual concentrations with an overall spatial influencing magnitude of 97.85%. Population, GDP, terrain, and LULC have individual spatial influencing magnitudes of 47.12% and 36.13%, 50.07% and 40.91% on PM2.5 annual concentrations respectively. Terrain and LULC are the dominating spatial influencing factors, and only these two factors together may approximately determine the spatial pattern of PM2.5 annual concentration over China with a high spatial influencing magnitude of 96.65%.

M-SAC-VLADNet: A Multi-Path Deep Feature Coding Model for Visual Classification

Vector of locally aggregated descriptor (VLAD) coding has become an efficient feature coding model for retrieval and classification. In some recent works, the VLAD coding method is extended to a deep feature coding model which is called NetVLAD. NetVLAD improves significantly over the original VLAD method. Although the NetVLAD model has shown its potential for retrieval and classification, the discriminative ability is not fully researched. In this paper, we propose a new end-to-end feature coding network which is more discriminative than the NetVLAD model. First, we propose a sparsely-adaptive and covariance VLAD model. Next, we derive the back propagation models of all the proposed layers and extend the proposed feature coding model to an end-to-end neural network. Finally, we construct a multi-path feature coding network which aggregates multiple newly-designed feature coding networks for visual classification. Some experimental results show that our feature coding network is very effective for visual classification.

Super-Resolution for "Jilin-1" Satellite Video Imagery via a Convolutional Network

Super-resolution for satellite video attaches much significance to earth observation accuracy, and the special imaging and transmission conditions on the video satellite pose great challenges to this task. The existing deep convolutional neural-network-based methods require pre-processing or post-processing to be adapted to a high-resolution size or pixel format, leading to reduced performance and extra complexity. To this end, this paper proposes a five-layer end-to-end network structure without any pre-processing and post-processing, but imposes a reshape or deconvolution layer at the end of the network to retain the distribution of ground objects within the image. Meanwhile, we formulate a joint loss function by combining the output and high-dimensional features of a non-linear mapping network to precisely learn the desirable mapping relationship between low-resolution images and their high-resolution counterparts. Also, we use satellite video data itself as a training set, which favors consistency between training and testing images and promotes the method’s practicality. Experimental results on “Jilin-1” satellite video imagery show that this method demonstrates a superior performance in terms of both visual effects and measure metrics over competing methods.

Skin Lesion Analysis towards Melanoma Detection Using Deep Learning Network

Skin lesions are a severe disease globally. Early detection of melanoma in dermoscopy images significantly increases the survival rate. However, the accurate recognition of melanoma is extremely challenging due to the following reasons: low contrast between lesions and skin, visual similarity between melanoma and non-melanoma lesions, etc. Hence, reliable automatic detection of skin tumors is very useful to increase the accuracy and efficiency of pathologists. In this paper, we proposed two deep learning methods to address three main tasks emerging in the area of skin lesion image processing, i.e., lesion segmentation (task 1), lesion dermoscopic feature extraction (task 2) and lesion classification (task 3). A deep learning framework consisting of two fully convolutional residual networks (FCRN) is proposed to simultaneously produce the segmentation result and the coarse classification result. A lesion index calculation unit (LICU) is developed to refine the coarse classification results by calculating the distance heat-map. A straight-forward CNN is proposed for the dermoscopic feature extraction task. The proposed deep learning frameworks were evaluated on the ISIC 2017 dataset. Experimental results show the promising accuracies of our frameworks, i.e., 0.753 for task 1, 0.848 for task 2 and 0.912 for task 3 were achieved.