RESUMEN

Drought is an extended shortage of rainfall resulting in water scarcity and affecting a region's social and economic conditions through environmental deterioration. Its adverse environmental effects can be minimised by timely prediction. Drought detection uses only ground observation stations, but satellite-based supervision scans huge land mass stretches and offers highly effective monitoring. This paper puts forward a novel drought monitoring system using satellite imagery by considering the effects of droughts that devastated agriculture in Thanjavur district, Tamil Nadu, between 2000 and 2022. The proposed method uses Holt Winter Conventional 2D-Long Short-Term Memory (HW-Conv2DLSTM) to forecast meteorological and agricultural droughts. It employs Climate Hazards Group InfraRed Precipitation with Station (CHIRPS) data precipitation index datasets, MODIS 11A1 temperature index, and MODIS 13Q1 vegetation index. It extracts the time series data from satellite images using trend and seasonal patterns and smoothens them using Holt Winter alpha, beta, and gamma parameters. Finally, an effective drought prediction procedure is developed using Conv2D-LSTM to calculate the spatiotemporal correlation amongst drought indices. The HW-Conv2DLSTM offers a better R2 value of 0.97. It holds promise as an effective computer-assisted strategy to predict droughts and maintain agricultural productivity, which is vital to feed the ever-increasing human population.

Asunto(s)

RESUMEN

Customer Relationship Management (CRM) is vital in modern business, aiding in the management and analysis of customer interactions. However, existing methods struggle to capture the dynamic and complex nature of customer relationships, as traditional approaches fail to leverage time series data effectively. To address this, we propose a novel GWO-attention-ConvLSTM model, which offers more effective prediction of customer churn and analysis of customer satisfaction. This model utilizes an attention mechanism to focus on key information and integrates a ConvLSTM layer to capture spatiotemporal features, effectively modeling complex temporal patterns in customer data. We validate our proposed model on multiple real-world datasets, including the BigML Telco Churn dataset, IBM Telco dataset, Cell2Cell dataset, and Orange Telecom dataset. Experimental results demonstrate significant performance improvements of our model compared to existing baseline models across these datasets. For instance, on the BigML Telco Churn dataset, our model achieves an accuracy of 95.17%, a recall of 93.66%, an F1 score of 92.89%, and an AUC of 95.00%. Similar results are validated on other datasets. In conclusion, our proposed GWO-attention-ConvLSTM model makes significant advancements in the CRM domain, providing powerful tools for predicting customer churn and analyzing customer satisfaction. By addressing the limitations of existing methods and leveraging the capabilities of deep learning, attention mechanisms, and optimization algorithms, our model paves the way for improving customer relationship management practices and driving business success.

RESUMEN

Purpose: To enhance the accuracy of real-time four-dimensional cone beam CT (4D-CBCT) imaging by incorporating spatiotemporal correlation from the sequential projection image into the single projection-based 4D-CBCT estimation process. Methods: We first derived 4D deformation vector fields (DVFs) from patient 4D-CT. Principal component analysis (PCA) was then employed to extract distinctive feature labels for each DVF, focusing on the first three PCA coefficients. To simulate a wide range of respiratory motion, we expanded the motion amplitude and used random sampling to generate approximately 900 sets of PCA labels. These labels were used to produce 900 simulated 4D-DVFs, which in turn deformed the 0% phase 4D-CT to obtain 900 CBCT volumes with continuous motion amplitudes. Following this, the forward projection was performed at one angle to get all of the digital reconstructed radiographs (DRRs). These DRRs and the PCA labels were used as the training data set. To capture the spatiotemporal correlation in the projections, we propose to use the convolutional LSTM (ConvLSTM) network for PCA coefficient estimation. For network testing, when several online CBCT projections (with different motion amplitudes that cover the full respiration range) are acquired and sent into the network, the corresponding 4D-PCA coefficients will be obtained and finally lead to a full online 4D-CBCT prediction. A phantom experiment is first performed with the XCAT phantom; then, a pilot clinical evaluation is further conducted. Results: Results on the XCAT phantom and the patient data show that the proposed approach outperformed other networks in terms of visual inspection and quantitative metrics. For the XCAT phantom experiment, ConvLSTM achieves the highest quantification accuracy with MAPE(Mean Absolute Percentage Error), PSNR (Peak Signal-to-Noise Ratio), and RMSE(Root Mean Squared Error) of 0.0459, 64.6742, and 0.0011, respectively. For the patient pilot clinical experiment, ConvLSTM also achieves the best quantification accuracy with that of 0.0934, 63.7294, and 0.0019, respectively. The quantification evaluation labels that we used are 1) the Mean Absolute Error (MAE), 2) the Normalized Cross Correlation (NCC), 3)the Structural Similarity Index Measurement(SSIM), 4)the Peak Signal-to-Noise Ratio (PSNR), 5)the Root Mean Squared Error(RMSE), and 6) the Absolute Percentage Error (MAPE). Conclusion: The spatiotemporal correlation-based respiration motion modeling supplied a potential solution for accurate real-time 4D-CBCT reconstruction.

RESUMEN

In the global challenge of Coronavirus disease 2019 (COVID-19) pandemic, accurate prediction of daily new cases is crucial for epidemic prevention and socioeconomic planning. In contrast to traditional local, one-dimensional time-series data-based infection models, the study introduces an innovative approach by formulating the short-term prediction problem of new cases in a region as multidimensional, gridded time series for both input and prediction targets. A spatial-temporal depth prediction model for COVID-19 (ConvLSTM) is presented, and further ConvLSTM by integrating historical meteorological factors (Meteor-ConvLSTM) is refined, considering the influence of meteorological factors on the propagation of COVID-19. The correlation between 10 meteorological factors and the dynamic progression of COVID-19 was evaluated, employing spatial analysis techniques (spatial autocorrelation analysis, trend surface analysis, etc.) to describe the spatial and temporal characteristics of the epidemic. Leveraging the original ConvLSTM, an artificial neural network layer is introduced to learn how meteorological factors impact the infection spread, providing a 5-day forecast at a 0.01° × 0.01° pixel resolution. Simulation results using real dataset from the 3.15 outbreak in Shanghai demonstrate the efficacy of Meteor-ConvLSTM, with reduced RMSE of 0.110 and increased R 2 of 0.125 (original ConvLSTM: RMSE = 0.702, R 2 = 0.567; Meteor-ConvLSTM: RMSE = 0.592, R 2 = 0.692), showcasing its utility for investigating the epidemiological characteristics, transmission dynamics, and epidemic development.

RESUMEN

Characterization of lung sounds (LS) is indispensable for diagnosing respiratory pathology. Although conventional neural networks (NNs) have been widely employed for the automatic diagnosis of lung sounds, deep neural networks can potentially be more useful than conventional NNs by allowing accurate classification without requiring preprocessing and feature extraction. Utilizing the long short-term memory (LSTM) layers to reveal the sequence-based properties of the LS time series, a novel architecture consisting of a cascade of convolutional long short-term memory (ConvLSTM) and LSTM layers, namely ConvLSNet is developed, which permits highly accurate diagnosis of pulmonary disease states. By modeling the multichannel lung sounds through the ConvLSTM layer, the proposed ConvLSNet architecture can concurrently deal with the spatial and temporal properties of the six-channel LS recordings without heavy preprocessing or data transformation. Notably, the proposed model achieves a classification accuracy of 97.4 % based on LS data corresponding to three pulmonary conditions, namely asthma, COPD, and the healthy state. Compared with architectures consisting exclusively of CNN or LSTM layers, as well as those employing a cascade integration of 2DCNN and LSTM layers, the proposed ConvLSNet architecture exhibited the highest classification accuracy, while imposing the lowest computational cost as quantified by the number of parameters, training time, and learning rate.

Asunto(s)

RESUMEN

Short-term precipitation forecasting methods are mainly divided into statistical forecasting, numerical model-based forecasting, and radar image extrapolation techniques. The two methods based on statistical prediction and numerical model have the disadvantages of being unstable and generating large errors. Therefore, this study proposes the use of deep learning for radar image extrapolation for precipitation forecasting, in particular by developing algorithms for ConvLSTM and SmaAT-UNet. The ConvLSTM model is a fusion of a CNN (Convolutional Neural Network) and LSTM (Long Short-Term Memory network), which solves the challenge of processing spatial sequence data, which is a task that traditional LSTM models cannot accomplish. At the same time, SmaAT-UNet enhances the traditional UNet structure by incorporating the CBAM (Convolutional Block Attention Module) attention mechanism and replacing the standard convolutional layer with depthwise separable convolution. This innovative approach aims to improve the efficiency and accuracy of short-term precipitation forecasting by improving feature extraction and data processing techniques. Evaluation and analysis of experimental data show that both models exhibit good predictive ability, with the SmaAT-UNet model outperforming ConvLSTM in terms of accuracy. The results show that the performance indicators of precipitation prediction, especially detection probability (POD) and the Critical Success index (CSI), show a downward trend with the extension of the prediction time. This trend highlights the inherent challenges of maintaining predictive accuracy over longer periods of time and highlights the superior performance and resilience of the SmaAT-UNet model under these conditions. Compared with the statistical forecasting method and numerical model forecasting method, its accuracy in short-term rainfall forecasting is improved.

RESUMEN

High speed side-view videos of sliding drops enable researchers to investigate drop dynamics and surface properties. However, understanding the physics of sliding requires knowledge of the drop width. A front-view perspective of the drop is necessary. In particular, the drop's width is a crucial parameter owing to its association with the friction force. Incorporating extra cameras or mirrors to monitor changes in the width of drops from a front-view perspective is cumbersome and limits the viewing area. This limitation impedes a comprehensive analysis of sliding drops, especially when they interact with surface defects. Our study explores the use of various regression and multivariate sequence analysis (MSA) models to estimate the drop width at a solid surface solely from side-view videos. This approach eliminates the need to incorporate additional equipment into the experimental setup. In addition, it ensures an unlimited viewing area of sliding drops. The Long Short Term Memory (LSTM) model with a 20 sliding window size has the best performance with the lowest root mean square error (RMSE) of 67 µm. Within the spectrum of drop widths in our dataset, ranging from 1.6 to 4.4 mm, this RMSE indicates that we can predict the width of sliding drops with an error of 2.4%. Furthermore, the applied LSTM model provides a drop width across the whole sliding length of 5 cm, previously unattainable.

RESUMEN

Epilepsy is a disease caused by abnormal neural discharge, which severely harms the health of patients. Its pathogenesis is complex and variable with various forms of seizures, leading to significant differences in epilepsy manifestations among different patients. The changes of brain network are strongly correlated with related pathologies. Therefore, it is crucial to effectively and deeply explore the intrinsic features of epilepsy signals to reveal the rules of epilepsy occurrence and achieve accurate detection. Existing methods have faced the following issues: 1) single approach for feature extraction, resulting in insufficient classification information due to the lack of rich dimensions in captured features; 2) inability to deeply analyze the essential commonality of epilepsy signal after feature extraction, making the model susceptible to data distribution and noise interference. Thus, we proposed a high-precision and robust model for epileptic seizure detection, which, for the first time, applies hypergraph convolution to the field of epilepsy detection. Through a hypergraph network structure constructed based on relationships between channels in electroencephalogram (EEG) signals, the model explores higher-order characteristics of epilepsy EEG data. Specifically, we use the Conv-LSTM module and Power spectral density (PSD), a two-branch parallel method, to extract channel features from space-time and frequency domains to solve the problem of insufficient feature extraction, and can adequately describe the data structure and distribution from multiple perspectives through double-branch parallel feature extraction. In addition, we construct a hypergraph on the captured features to explore the intrinsic features in the high-dimensional space in an attempt to reveal the essential commonality of epileptic signal feature extraction. Finally, using the ensemble learning concept, we accomplished epilepsy detection on the dual-branch hypergraph convolution. The model underwent leave-one-out cross-validation on the TUH dataset, achieving an average accuracy of 96.9%, F1 score of 97.3%, Pre of 98.2% and Re of 96.7%. In addition, the model was generalized performance tested on CHB-MIT scalp EEG dataset with leave-one-out cross-validation, and the average ACC, F1 score, Pre and Re were 94.4%, 95.1%, 95.8%, and 93.9% respectively. Experimental results indicate that the model outperforms related literature, providing valuable reference for the clinical application of epilepsy detection.

RESUMEN

Protected areas (PAs) are effective in mitigating human pressures, yet their future pressure alleviating effects remain unclear. In this study, we employed the ConvLSTM model to forecast the future human footprint and analyzed human pressure trends using Theil-Sen median and Mann-Kendall tests. We further evaluated the mitigating effects of PAs within their buffer zones (1-10 km) and the contributions of different IUCN categories of PAs to mitigating human pressure using linear regression models. The results indicate that by 2035, the average human pressure value is expected to increase by 11%, with trends exhibiting a polarized pattern. Furthermore, PAs also effectively mitigate human pressure within their 1 km buffer zones. Different categories of PAs vary in their effectiveness in mitigating human pressure, and stricter conservation areas are not always the most effective. This study can offer insights for evaluating the effectiveness of PAs in reducing human pressure and advocate for their targeted management in urban areas.

Asunto(s)

RESUMEN

Spectrum prediction is a promising technique to release spectrum resources and plays an essential role in cognitive radio networks and spectrum situation generating. Traditional algorithms normally focus on one-dimensional or predict spectrum values in a slot-by-slot manner and thus cannot fully perceive the spectrum states in complex environments and lack timeliness. In this paper, a deep learning-based prediction method with a simple structure is developed for temporal-spectral and multi-slot spectrum prediction simultaneously. Specifically, we first analyze and construct spectrum data suitable for the model to simultaneously achieve long-term and multi-dimensional spectrum prediction. Then, a hierarchical spectrum prediction system is developed that takes advantage of the advanced Bi-ConvLSTM and the seq2seq framework. The Bi-ConvLSTM captures time-frequency characteristics of spectrum data, and the seq2seq framework is used for long-term spectrum prediction. Furthermore, the attention mechanism is used to address the limitations of the seq2seq framework that compresses all inputs into fixed-length vectors, resulting in information loss. Finally, the experimental results have shown that the proposed model has a significant advantage over the benchmark schemes. Especially, the proposed spectrum prediction model achieves 6.15%, 0.7749, 1.0978, and 0.9628 in MAPE, MAE, RMSE, and R2, respectively, which is better than all the baseline deep learning models.

RESUMEN

The Vegetation Health Index (VHI) is a metric used to assess the health and condition of vegetation, based on satellite-derived data. It offers a comprehensive indicator of stress or vigor, commonly used in agriculture, ecology, and environmental monitoring for forecasting changes in vegetation health. Despite its advantages, there are few studies on forecasting VHI as a future projection, particularly using up-to-date and effective machine learning methods. Hence, the primary objective of this study is to forecast VHI values by utilizing remotely sensed images. To achieve this objective, the study proposes employing a combined Convolutional Neural Network (CNN) and a specific type of Recurrent Neural Network (RNN) called Long Short-Term Memory (LSTM), known as ConvLSTM. The VHI time series images are calculated based on the Normalized Difference Vegetation Index (NDVI) and Land Surface Temperature (LST) data obtained from the Moderate Resolution Imaging Spectroradiometer (MODIS) aboard the Terra and Aqua satellites. In addition to the traditional image-based calculation, the study suggests using global minimum and global maximum values (global scale) of NDVI and LST time series for calculating the VHI. The results of the study showed that the ConvLSTM with a 1-layer structure generally provided better forecasts than 2-layer and 3-layer structures. The average Root Mean Square Error (RMSE) values for the 1-step, 2-step, and 3-step ahead VHI forecasts were 0.025, 0.026, and 0.026, respectively, with each step representing an 8-day forecast horizon. Moreover, the proposed global scale model using the applied ConvLSTM structures outperformed the traditional VHI calculation method.

Asunto(s)

RESUMEN

Sparse view computed tomography (SVCT) aims to reduce the number of X-ray projection views required for reconstructing the cross-sectional image of an object. While SVCT significantly reduces X-ray radiation dose and speeds up scanning, insufficient projection data give rise to issues such as severe streak artifacts and blurring in reconstructed images, thereby impacting the diagnostic accuracy of CT detection. To address this challenge, a dual-domain reconstruction network incorporating multi-level wavelet transform and recurrent convolution is proposed in this paper. The dual-domain network is composed of a sinogram domain network (SDN) and an image domain network (IDN). Multi-level wavelet transform is employed in both IDN and SDN to decompose sinograms and CT images into distinct frequency components, which are then processed through separate network branches to recover detailed information within their respective frequency bands. To capture global textures, artifacts, and shallow features in sinograms and CT images, a recurrent convolution unit (RCU) based on convolutional long and short-term memory (Conv-LSTM) is designed, which can model their long-range dependencies through recurrent calculation. Additionally, a self-attention-based multi-level frequency feature normalization fusion (MFNF) block is proposed to assist in recovering high-frequency components by aggregating low-frequency components. Finally, an edge loss function based on the Laplacian of Gaussian (LoG) is designed as the regularization term for enhancing the recovery of high-frequency edge structures. The experimental results demonstrate the effectiveness of our approach in reducing artifacts and enhancing the reconstruction of intricate structural details across various sparse views and noise levels. Our method excels in both performance and robustness, as evidenced by its superior outcomes in numerous qualitative and quantitative assessments, surpassing contemporary state-of-the-art CNNs or Transformer-based reconstruction methods.

Asunto(s)

RESUMEN

Nanomaterial-based aptasensors serve as useful instruments for detecting small biological entities. This work utilizes data gathered from three electrochemical aptamer-based sensors varying in receptors, analytes of interest, and lengths of signals. Our ultimate objective was the automatic detection and quantification of target analytes from a segment of the signal recorded by these sensors. Initially, we proposed a data augmentation method using conditional variational autoencoders to address data scarcity. Secondly, we employed recurrent-based networks for signal extrapolation, ensuring uniform signal lengths. In the third step, we developed seven deep learning classification models (GRU, unidirectional LSTM (ULSTM), bidirectional LSTM (BLSTM), ConvGRU, ConvULSTM, ConvBLSTM, and CNN) to identify and quantify specific analyte concentrations for six distinct classes, ranging from the absence of analyte to 10 µM. Finally, the second classification model was created to distinguish between abnormal and normal data segments, detect the presence or absence of analytes in the sample, and, if detected, identify the specific analyte and quantify its concentration. Evaluating the time series forecasting showed that the GRU-based network outperformed two other ULSTM and BLSTM networks. Regarding classification models, it turned out signal extrapolation was not effective in improving the classification performance. Comparing the role of the network architectures in classification performance, the result showed that hybrid networks, including both convolutional and recurrent layers and CNN networks, achieved 82% to 99% accuracy across all three datasets. Utilizing short-term Fourier transform (STFT) as the preprocessing technique improved the performance of all datasets with accuracies from 84% to 99%. These findings underscore the effectiveness of suitable data preprocessing methods in enhancing neural network performance, enabling automatic analyte identification and quantification from electrochemical aptasensor signals.

RESUMEN

Accurate and reliable prediction of air pollutant concentrations is important for rational avoidance of air pollution events and government policy responses. However, due to the mobility and dynamics of pollution sources, meteorological conditions, and transformation processes, pollutant concentration predictions are characterized by great uncertainty and instability, making it difficult for existing prediction models to effectively extract spatial and temporal correlations. In this paper, a powerful pollutant prediction model (STA-ResConvLSTM) is proposed to achieve accurate prediction of pollutant concentrations. The model consists of a deep learning network model based on a residual neural network (ResNet), a spatial-temporal attention mechanism, and a convolutional long short-term memory neural network (ConvLSTM). The spatial-temporal attention mechanism is embedded in each residual unit of the ResNet to form a new residual neural network with the spatial-temporal attention mechanism (STA-ResNet). Deep extraction of spatial-temporal distribution features of pollutant concentrations and meteorological data from several cities is carried out using STA-ResNet. Its output is used as an input to the ConvLSTM, which is further analyzed to extract preliminary spatial-temporal distribution features extracted from the STA-ResNet. The model realizes the spatial-temporal correlation of the extracted feature sequences to accurately predict pollutant concentrations in the future. In addition, experimental studies on urban agglomerations around Long Beijing show that the prediction model outperforms various popular baseline models in terms of accuracy and stability. For the single-step prediction task, the proposed pollutant concentration prediction model performs well, exhibiting a root-mean-square error (RMSE) of 9.82. Furthermore, even for the pollutant prediction task of 1 to 48 h, we performed a multi-step prediction and achieved a satisfactory performance, being able to achieve an average RMSE value of 13.49.

RESUMEN

Falls represent a significant health concern for the elderly. While studies on deep learning-based preimpact fall detection have been conducted to mitigate fall-related injuries, additional efforts are needed for embedding in microcomputer units (MCUs). In this study, ConvLSTM, the state-of-the-art model, was benchmarked, and we attempted to lightweight it by leveraging features from image-classification models VGGNet and ResNet while maintaining performance for wearable airbags. The models were developed and evaluated using data from young subjects in the KFall public dataset based on an inertial measurement unit (IMU), leading to the proposal of TinyFallNet based on ResNet. Despite exhibiting higher accuracy (97.37% < 98.00%) than the benchmarked ConvLSTM, the proposed model requires lower memory (1.58 MB > 0.70 MB). Additionally, data on the elderly from the fall data of the FARSEEING dataset and activities of daily living (ADLs) data of the KFall dataset were analyzed for algorithm validation. This study demonstrated the applicability of image-classification models to preimpact fall detection using IMU and showed that additional tuning for lightweighting is possible due to the different data types. This research is expected to contribute to the lightweighting of deep learning models based on IMU and the development of applications based on IMU data.

Asunto(s)

RESUMEN

Precipitation nowcasting refers to the use of specific meteorological elements to predict precipitation in the next 0-2 h. Existing methods use radar echo maps and the Z-R relationship to directly predict future rainfall rates through deep learning methods, which are not physically constrained, but suffer from severe loss of predicted image details. This paper proposes a new model framework to effectively solve this problem, namely LSTMAtU-Net. It is based on the U-Net architecture, equipped with a Convolutional LSTM (ConvLSTM) unit with the vertical flow direction and depthwise-separable convolution, and we propose a new component, the Efficient Channel and Space Attention (ECSA) module. The ConvLSTM unit with the vertical flow direction memorizes temporal changes by extracting features from different levels of the convolutional layers, while the ECSA module innovatively integrates different structural information of each layer of U-Net into the channelwise attention mechanism to learn channel and spatial information, thereby enhancing attention to the details of precipitation images. The experimental results showed that the performance of the model on the test dataset was better than other examined models and improved the accuracy of medium- and high-intensity precipitation nowcasting.

Asunto(s)

RESUMEN

The three-dimensional convolutional neural network (3D-CNN) and long short-term memory (LSTM) have consistently outperformed many approaches in video-based facial expression recognition (VFER). The image is unrolled to a one-dimensional vector by the vanilla version of the fully-connected LSTM (FC-LSTM), which leads to the loss of crucial spatial information. Convolutional LSTM (ConvLSTM) overcomes this limitation by performing LSTM operations in convolutions without unrolling, thus retaining useful spatial information. Motivated by this, in this paper, we propose a neural network architecture that consists of a blend of 3D-CNN and ConvLSTM for VFER. The proposed hybrid architecture captures spatiotemporal information from the video sequences of emotions and attains competitive accuracy on three FER datasets open to the public, namely the SAVEE, CK + , and AFEW. The experimental results demonstrate excellent performance without external emotional data with the added advantage of having a simple model with fewer parameters. Moreover, unlike the state-of-the-art deep learning models, our designed FER pipeline improves execution speed by many factors while achieving competitive recognition accuracy. Hence, the proposed FER pipeline is an appropriate candidate for recognizing facial expressions on resource-limited embedded platforms for real-time applications.

RESUMEN

Currently, three-dimensional convolutional neural networks (3DCNNs) are a popular approach in the field of human activity recognition. However, due to the variety of methods used for human activity recognition, we propose a new deep-learning model in this paper. The main objective of our work is to optimize the traditional 3DCNN and propose a new model that combines 3DCNN with Convolutional Long Short-Term Memory (ConvLSTM) layers. Our experimental results, which were obtained using the LoDVP Abnormal Activities dataset, UCF50 dataset, and MOD20 dataset, demonstrate the superiority of the 3DCNN + ConvLSTM combination for recognizing human activities. Furthermore, our proposed model is well-suited for real-time human activity recognition applications and can be further enhanced by incorporating additional sensor data. To provide a comprehensive comparison of our proposed 3DCNN + ConvLSTM architecture, we compared our experimental results on these datasets. We achieved a precision of 89.12% when using the LoDVP Abnormal Activities dataset. Meanwhile, the precision we obtained using the modified UCF50 dataset (UCF50mini) and MOD20 dataset was 83.89% and 87.76%, respectively. Overall, our work demonstrates that the combination of 3DCNN and ConvLSTM layers can improve the accuracy of human activity recognition tasks, and our proposed model shows promise for real-time applications.

Asunto(s)

RESUMEN

Breast ultrasound (BUS) imaging has become one of the key imaging modalities for medical image diagnosis and prognosis. However, the manual process of lesion delineation from ultrasound images can incur various challenges concerning variable shape, size, intensity, curvature, or other medical priors of the lesion in the image. Therefore, computer-aided diagnostic (CADx) techniques incorporating deep learning-based neural networks are automatically used to segment the lesion from BUS images. This paper proposes an encoder-decoder-based architecture to recognize and accurately segment the lesion from two-dimensional BUS images. The architecture is utilized with the residual connection in both encoder and decoder paths; bi-directional ConvLSTM (BConvLSTM) units in the decoder extract the minute and detailed region of interest (ROI) information. BConvLSTM units and residual blocks help the network weigh ROI information more than the similar background region. Two public BUS image datasets, one with 163 images and the other with 42 images, are used. The proposed model is trained with the augmented images (ten forms) of dataset one (with 163 images), and test results are produced on the second dataset and the testing set of the first dataset-the segmentation performance yielding comparable results with the state-of-the-art segmentation methodologies. Similarly, the visual results show that the proposed approach for BUS image segmentation can accurately identify lesion contours and can potentially be applied for similar and larger datasets.

Asunto(s)

RESUMEN

Background: High precision segmentation of retinal blood vessels from retinal images is a significant step for doctors to diagnose many diseases such as glaucoma and cardiovascular diseases. However, at the peripheral region of vessels, previous U-Net-based segmentation methods failed to significantly preserve the low-contrast tiny vessels. Methods: For solving this challenge, we propose a novel network model called Bi-directional ConvLSTM Residual U-Net (BCR-UNet), which takes full advantage of U-Net, Dropblock, Residual convolution and Bi-directional ConvLSTM (BConvLSTM). In this proposed BCR-UNet model, we propose a novel Structured Dropout Residual Block (SDRB) instead of using the original U-Net convolutional block, to construct our network skeleton for improving the robustness of the network. Furthermore, to improve the discriminative ability of the network and preserve more original semantic information of tiny vessels, we adopt BConvLSTM to integrate the feature maps captured from the first residual block and the last up-convolutional layer in a nonlinear manner. Results and discussion: We conduct experiments on four public retinal blood vessel datasets, and the results show that the proposed BCR-UNet can preserve more tiny blood vessels at the low-contrast peripheral regions, even outperforming previous state-of-the-art methods.

Asunto(s)