End-stage renal disease (ESRD) is a severe kidney disorder; kidney ultrasound, as a non invasive diagnostic tool, is widely used in its clinical diagnosis. However, due to the morphological changes caused by ESRD, such as kidney shrinkage, cortical thinning, and increased echogenicity, traditional ultrasound image segmentation and diagnosis still face significant challenges. To improve segmentation accuracy and diagnostic performance, this paper proposes a multi-task learning-based approach for ultrasound kidney segmentation and auxiliary diagnosis. The method combines the classical UNet architecture with the advanced RKAN-ResNet34 encoder and incorporates an improved Convolutional Block Attention Module (CBAM), which integrates edge attention and deformable convolutions, to address the issues of blurred boundaries and morphological changes in kidney images. By jointly optimizing segmentation and classification tasks, the model simultaneously enhances both kidney segmentation accuracy and auxiliary diagnostic performance. Experimental results show that the proposed method achieves a Dice coefficient of 0.9233 and an IoU of 0.8592 for segmentation, along with an accuracy of 98.64% and an F1 score of 0.9882 for auxiliary diagnosis, outperforming existing methods. This study provides an effective solution for the automation of the ultrasound kidney image segmentation and diagnosis, contributing to the auxiliary diagnosis of end-stage renal disease.

Segmenting fetal heart ultrasound images presents significant challenges due to variations in image quality, noise, and the small size of the target region. This paper introduces the BRM-UNet framework, which incorporates multiscale feature learning, boundary refinement, and a selective scanning mechanism to overcome the limitations of current methods. The proposed model utilizes a weighted loss function that combines Dice and boundary-aware losses, enhancing both segmentation accuracy and boundary precision. Moreover, the selective scanning mechanism improves feature extraction by prioritizing the heart region, ensuring more accurate segmentation. Experimental results on the FOCUS dataset show that our model outperforms existing approaches, achieving superior segmentation performance (Dice coefficient 0.9178, IoU 0.8500) and boundary precision (Hausdorff distance 27.2503). Further validation on a pediatric cardiac ultrasound dataset yielded excellent results (Left atrium Dice 0.9000, Left ventricle Dice 0.9035), demonstrating the model's robust generalization and its potential as a more accurate and automated solution for both fetal and pediatric cardiac ultrasound assessments.

To avoid severe malfunctions of industrial equipment, it is necessary to perform accurate detection in the early stages of abnormal occurrences. However, early anomalies are usually weak, difficult to model anomalous features, and affected by data uncertainty and training uncertainty. To address these limitations, we propose a multi-task and multi-domain temporal memory autoencoder (MTMAE). In the encoding stage, we design a temporal feature learning convolutional encoder and a frequency-aware temporal block to fuse time-domain and frequency-domain anomalous features, thereby creating a cloud-enhancement feature modeling (CEFM) approach based on cloud model theory to mitigate uncertainty in the data. After obtaining the latent features of the encoder, the reconstruction task uses the deconvolution network to recover the data, forming an encoder-decoder adaptive matching. The encoding memory task uses an external attention memory unit scorer to memorize potential patterns in the data. In addition, we design an optimized regularization uncertainty weighting (UW) method to balance the two tasks and penalize training uncertainty. The experimental results of five public datasets demonstrate the superiority of MTMAE in anomaly detection, with an average F1 score of 0.871 and an area under the precision-recall curve of 0.887. In the actual anomaly detection of private marine diesel engine data, MTMAE can detect early weak anomalies fastest and has the lowest false alarm rate. In addition, we also demonstrated the contribution of CEFM and UW methods to the model's resistance to uncertainty through noise set detection and model-independent detection experiments.

Gastric cancer, as a malignant tumor, is one of the most common cancer-related deaths worldwide with high mortality and incidence rates. Therefore, the endoscopic detection of gastric cancer at an early stage is essential. In this paper, we propose an automatic diagnosis method of early gastric cancer (EGC) based on deep learning (DL) techniques. Specifically, with the new annotated endoscopic image dataset collected from a single-center, this paper designs several DL architectures to realize the automatic analysis of EGC images. Particularly, a guided-attention deep network was introduced, derived from ResNet-50, for the accurate score prediction of EGC and the extraction of feature information. Furthermore, we combined a lightweight attention module and multiscale feature extractor with U-Net for estimating the pixel-level segmentation of EGC pathological regions. Experiment results on the presented dataset showed outstanding performance in the involved classification and segmentation assignments with an accuracy of 98.84% and an intersection over union (IOU) of 0.64, revealing the potential applications of DL in aiding and improving EGC diagnosis using endoscopic images.

Great changes have been brought about by the coastal environment when the economy develops rapidly. Coastal environmental monitoring is the basis and technical guarantee for coastal environmental protection supervision and management. It is one of the important tasks to detect and timely discover coastal seawater anomalies. Usually, a single sensor cannot determine whether the coastal environment or ship operation is an anomaly. Recently, an unmanned surface vehicle for coastal environment monitoring was developed, and stacked autoencoders are used for seawater anomaly detection using multisensor data fusion methods. The multisensor data of pH, conductivity, and ammonia nitrogen are employed to judge the anomaly of seawater. The mean, standard deviation, mean square root, and normalized power spectrum features of multisensor data are extracted, and a stacked autoencoder is employed to fuse these features for anomaly detection. The proposed method is feasible and effective for anomaly detection of coastal water quality and ship operation. Compared with other commonly used methods, the proposed method has a higher recall, precision, and F1 score performance.

Hydropower generation has become an important guarantee for social and economic evolution. The hydropower turbine is the main power device of hydropower generation. And the anomaly state of the turbine is important to the operation and maintenance of the turbine. The complex signals measured by the hydropower turbine sensor are decomposed into the sub signals employing variational modal decomposition (VMD). The accuracy of anomaly detection is improved by applying VMD. A hydropower turbine anomaly detection application combining VMD and hierarchical temporal memory (HTM) is developed. Specificities of spinning-tree, cluster, HTM and the proposed method are 0.730, 0.690, 0.820, 0.945, respectively. Accuracy are 0.872, 0.873, 0.931, 0.959, respectively, and F1 scores are 0.923, 0.924, 0.959, 0.976, respectively. The proposed method can be applied to other anomaly detection applications.

Unmanned surface vehicle (USV) is the future development direction of ships, but few studies have focused on USV’s energy optimization based on visual perception. An energy optimization strategy based on visual object detection is developed for USV. A visual target recognition method is proposed by combining YOLOv5 and DeepSORT. Visual recognition results are fused with radar targets to support route plan for energy optimization of USV. By dynamically adjusting the threshold of visual target recognition with the target number provided by radar, the target detection result is more accurate. Experimental results show that the proposed target detection method has the best performance than other commonly used methods, MOTA of the proposed method reaches 87.40%, and the YOLOv4 method, CenterTrack and FairMOT are 85.18%, 64.97% and 46.39% respectively. And the energy consumption optimization can be dynamically achieved by continuously analyzing the speed and path of the USV and predicting fuel consumption.

Gastroscopy is the preferred method to detect upper gastrointestinal lesions and has been widely adopted. For the diagnosis of gastrointestinal diseases, the first crucial step is to properly recognize the anatomical location. Image recognition using deep learning algorithms has made remarkable progress in the medical fields, and been increasingly used in gastrointestinal endoscopy. However, due to the similarity of many parts of the gastrointestinal tract, the accuracy of multi-site recognition based on traditional convolution neural network still needs to be improved. The effectiveness of convolutional neural networks (CNNs) to recognize anatomical sites in endoscopic images is explored and a channel-separation strategy for classification based on richer convolutional features is proposed. We use 1 x 1 conv. layer with channel depth 12 to extract and separate the depth feature information of anatomical sites and use global average pooling to output the final results. Compared with several classical networks, the proposed method is feasible and can improve the performance of anatomical site recognition of endoscopic images. The proposed method achieved 98.84% accuracy, 92.86% precision and 92.43% F1 score. The experimental results show that the proposed method has the potential to be applied to the automatic recognition of anatomical sites for clinical gastroscopy.

Objective:To evaluate the diagnostic value of the convolution neural network model DenseNet121 based on deep learning in the diagnosis of end-stage renal disease (ESRD).Methods:In this retrospective study, 489 kidney ultrasound images of patients diagnosed with end-stage renal disease from January 1, 2019 to September 30, 2019 and 450 kidney ultrasound images of healthy controls were selected at China-Japan Friendship Hospital. The deep learning-based supervised convolutional neural network model DenseNet121 was used for network training and verification. According to whether it was end-stage renal disease or not, the prediction results of the deep learning-based model were compared with the prediction results of professional imaging physicians. Receiver operating characteristic (ROC) curve analysis was used to evaluate the performance of the deep learning-based model, the accuracy, specificity, sensitivity, and area under the curve (AUC) were used as metrics to compare the performance of the deep learning-based model and professional imaging physicians, and Delong was used to compare the difference of AUC.Results:The prediction accuracy of professional imaging physicians for end-stage renal disease was 89.36%, the sensitivity was 81.63%, the specificity was 97.77%, and the AUC was 0.897. The prediction accuracy of the deep learning-based convolution neural network model DenseNet121 for end-stage renal disease was 93.51%, the sensitivity was 96.12%, the specificity was 90.66%, and the AUC was 0.934. Compared with professional physicians, the DenseNet121 model had higher diagnostic ability (Z=3.034, P=0.002).Conclusion:The ultrasonic diagnosis method based on deep learning shows high diagnostic performance, and it has the potential to assist professional imaging physicians in the diagnosis of end-stage renal disease.

With the development of unmanned systems, more and more attentions are paid to the energy and power systems of data-driven ships. The autonomy of unmanned ships puts forward urgent requirements for the monitoring and prediction of the energy and power system of ships. Aiming at the state prediction for marine diesel engine, an improvement method based on variational modal decomposition (VMD) and long short-term memory (LSTM) is proposed in this paper. The sub signals are obtained by decomposing the signal to be predicted through VMD, the sub signals and resident signal are all predicted with LSTM, and the reconstruction prediction signal is obtained by sum all the predicted sub signals and resident signal. Compared with LSTM, ESN, and SVR methods, the proposed method reduces the prediction errors significantly. Compared with LSTM, the RE errors of the two sensors are reduced by 49.79% and 56.32% respectively, and the RMSE errors are reduced by 34.65% and 27.71% respectively. The performance of this method is better than other methods, and it has sufficient accuracy performance for state prediction of marine diesel engine.

BACKGROUND:The incidence rate of renal disease is high, which can cause end-stage renal disease. Ultrasound is a commonly used imaging method, including conventional ultrasound, color ultrasound, elastography, etc. Machine learning is a potential method which has been widely used in clinical practices.OBJECTIVE:To compare the diagnostic performance of different ultrasonic image measurement parameters for kidney diseases, and to compare different machine learning methods with the human- reading method.METHODS:Ninety-four patients with pathologically diagnosed renal diseases and 109 normal controls were included in this study. The patients were examined by conventional ultrasound, color ultrasound and shear wave elasticity, respectively. Ultrasonic data were analyzed by Support vector machine (SVM), random forest (RF), K-nearest neighbor (KNN) and artificial neural network (ANN), respectively, and compared with the human-reading method.RESULTS:Only ultrasound elastography data have a diagnostic value for renal diseases. The accuracy of SVM, RF, KNN and ANN methods is 80.98%, 80.32%, 78.03% and 79.67%, respectively, while the accuracy of human-reading is 78.33%. In the data of machine learning ultrasound elastography, the elastic hardness parameters of the renal cortex are most important.CONCLUSION:Ultrasound elastography is of the highest diagnostic value in machine learning for nephropathy, the diagnostic efficiency of the machine learning method is slightly higher than that of the human-reading method, and the diagnostic ability of the SVM method is higher than other methods.

Hydropower has become an important energy supply method due to its advantages of pollution-free, renewable, and abundant resources. With the development of detection technology, methods that can detect anomaly without stopping the hydropower unit have gradually gained attention. An autoencoder based on one-dimensional convolutional neural network is proposed for anomaly detection of bearing condition monitoring system of the hydropower unit. The input sensor signals contain 7 features, and the anomaly is judged by the residual error between the reconstructed data and the original data. The proposed method is compared with one-class support vector machine method and autoencoder method. The results show that the accuracy, precision and recall of the proposed method are significantly better than other methods. F1 score, AUC and confusion matrix also show that the proposed method is optimal. The proposed method achieves a detection effect superior to the existing methods with the combination of autoencoder and convolutional neural network.

Anomaly detection for hydropower turbine unit is a requirement for the safety of hydropower system. An unsupervised anomaly detection method employing variational modal decomposition (VMD) and deep autoencoder is proposed. VMD is employed to the data collected by multiple sensors to obtain the sub signal of each data. These sub signals in each time-period constitute two-dimensional data. The autoencoder based on convolutional neural network is used to complete unsupervised learning, and the reconstruction residual of autoencoder is used for anomaly detection. The experimental results show that the deep autoencoder can increase the interval between abnormal and normal data distribution, and VMD can effectively reduce the number of samples in the overlapping area. Compared with traditional autoencoder method, the proposed method improves the recall, precision and F1 scores by 0.140, 0.205 and 0.175, respectively. The proposed method achieves better anomaly detection performance than other methods.

Autonomous navigation has become a trend in the development of ships, and the realization of automatic obstacle avoidance plays an important role. Radar is an indispensable sensor for ships, and it provide information on the location and distance of water surface obstacles. Based on the cluster analysis of radar point cloud data, the number of water surface obstacles is obtained. Use the information to adjust the output threshold of the obstacle detection network. Realize the detection and classification of obstacle targets. Considering the reduction of time complexity, the DBSCAN algorithm based on KD tree acceleration can be used, where the KD tree parameter selection determines the clustering effect. This paper proposes to use flower pollination optimization algorithm to determine the two parameters to ensure higher clustering accuracy. This method can find the appropriate optimal parameters in the parameter space, which can make the performance of the DBSCAN clustering algorithm based on KD tree acceleration the best.

目的 基于剪切波弹性成像(SWE)量化参数和卷积神经网络建立深度学习(DL)模型预测肾脏病变.方法 采集94例肾脏病变患者(病例组)和109名健康人(对照组)的肾脏超声SWE量化参数.利用卷积神经网络建立DL模型,比较DL模型和支持向量机、随机森林模型预测肾脏病变的敏感度、特异度、准确率和曲线下面积(AUC).结果 DL模型对预测肾脏病变的敏感度为90.48％,特异度为100％,准确率为95.12％,AUC为0.93;支持向量机模型的敏感度、特异度、准确率和AUC分别为80.74％、80.71％、80.98％、0.90,随机森林模型分别为82.22％、77.87％、80.33％和0.88.DL模型预测敏感度、特异度、准确率和AUC均高于支持向量机和随机森林模型,与支持向量机模型和随机森林模型预测肾脏病变差异均有统计学意义(P均＜0.05).结论 基于SWE量化参数和卷积神经网络的DL模型预测肾脏疾病性能良好,具有一定临床价值.

BACKGROUND:Noninvasively predicting kidney tubulointerstitial fibrosis is important because it's closely correlated with the development and prognosis of chronic kidney disease (CKD). Most studies of shear wave elastography (SWE) in CKD were limited to non-linear statistical dependencies and didn't fully consider variables' interactions. Therefore, support vector machine (SVM) of machine learning was used to assess the prediction value of SWE and traditional ultrasound techniques in kidney fibrosis. METHODS:We consecutively recruited 117 CKD patients with kidney biopsy. SWE, B-mode, color Doppler flow imaging ultrasound and hematological exams were performed on the day of kidney biopsy. Kidney tubulointerstitial fibrosis was graded by semi-quantification of Masson staining. The diagnostic performances were accessed by ROC analysis. RESULTS:Tubulointerstitial fibrosis area was significantly correlated with eGFR among CKD patients (R = 0.450, P < 0.001). AUC of SWE, combined with B-mode and blood flow ultrasound by SVM, was 0.8303 (sensitivity, 77.19%; specificity, 71.67%) for diagnosing tubulointerstitial fibrosis (>10%), higher than either traditional ultrasound, or SWE (AUC, 0.6735 [sensitivity, 67.74%; specificity, 65.45%]; 0.5391 [sensitivity, 55.56%; specificity, 53.33%] respectively. Delong test, p < 0.05); For diagnosing different grades of tubulointerstitial fibrosis, SWE combined with traditional ultrasound by SVM, had AUCs of 0.6429 for mild tubulointerstitial fibrosis (11%-25%), and 0.9431 for moderate to severe tubulointerstitial fibrosis (>50%), higher than other methods (Delong test, p < 0.05). CONCLUSION:SWE with SVM modeling could improve the diagnostic performance of traditional kidney ultrasound in predicting different kidney tubulointerstitial fibrosis grades among CKD patients.

In Marine diesel engine vibration monitoring system, a large number of high frequency data can be detected by high frequency sensors. Using data compression technology can effectively reduce the amount of data monitored by the ship vibration monitoring system and improve the communication efficiency of the system. In this paper, a lossless compression method combining variational mode decomposition (VMD) and run length encoding (RLE) for marine diesel high-frequency data is proposed. First, the signal is decomposed by VMD, and then the decomposed sub-signals and residual signals are encoded. In order to obtain better compression ratio, the compression method of RLE and Lempel-Ziv-Welch (LZW) is adopted. Experimental results show that this method can compress data effectively for high frequency signal of diesel engine morning and save network bandwidth.

An intelligent platform prototype is established for a coastal environment monitoring ship. LSTM and GBDT methods are developed for pH value and fuel consumption prediction in the intelligent platform. The results of applying the general prediction algorithms to actual environments’ data and marine diesel engine data are reported. GBDT has the best predictive results with the smallest error. SVM and SVR have similar prediction effects, while FNN has the largest error. As the prediction time increases, the error of LSTM becomes large. The ship intelligence platform can provide unified data support and general intelligent algorithms for data-driven applications, and it has the potential to be widely used in coastal environmental monitoring applications.