Performance Of U-Net Research Articles

Combined Unet and CNN image classification model for COVID disease detection using CXR/CT imaging

Accurate SARS-CoV-2 screening is made possible by automated Computer-Aided Diagnosis (CAD) which reduces the stress on healthcare systems. Since Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) is highly contagious, the transition chain can be broken through an early diagnosis by clinical knowledge and Artificial Intelligence (AI). Manual findings are time and labor-intensive. Even if Reverse Transcription-Polymerase Chain Reaction (RT-PCR) delivers quick findings, Chest X-ray (CXR) imaging is still a more trustworthy tool for disease classification and assessment. Several studies have been conducted using Deep Learning (DL) algorithms for COVID-19 detection. One of the biggest challenges in modernizing healthcare is extracting useful data from high-dimensional, heterogeneous, and complex biological data. Intending to introduce an automated COVID-19 diagnosis model, this paper develops a proficient optimization model that enhances the classification performance with better accuracy. The input images are initially pre-processed with an image filtering approach for noise removal and data augmentation to extend the dataset. Secondly, the images are segmented via U-Net and are given to classification using the Fused U-Net Convolutional Neural Network (FUCNN) model. Here, the performance of U-Net is enhanced through the modified Moth Flame Optimization (MFO) algorithm named Chaotic System-based MFO (CSMFO) by optimizing the weights of U-Net. The significance of the implemented model is confirmed over a comparative evaluation with the state-of-the-art models. Specifically, the proposed CSMFO-FUCNN attained 98.45% of accuracy, 98.63% of sensitivity, 98.98% of specificity, and 98.98% of precision.

Optimization and Performance Evaluation of Deep Learning Algorithm in Medical Image Processing

In this paper, the optimization and performance evaluation of deep learning algorithm in medical image processing are studied. Firstly, the paper introduces the importance and challenges of medical image processing, and expounds the application prospect of deep learning in this field. Subsequently, this paper discusses the optimization methods of deep learning algorithm in detail, including model structure design, data preprocessing, super parameter adjustment and so on. In terms of performance evaluation, this study selected classic models such as U-Net, DeepLab and DenseNet, and compared them with ROC curve and AUC value to evaluate their predictive ability in medical image classification. The results show that the DenseNet model shows high performance in prediction accuracy, while the performance of U-Net and DeepLab models is slightly average. Finally, the advantages and disadvantages of each model are analyzed, and the future research direction is prospected. This study is of great significance to promote the development and application of medical image processing technology, and provides important theoretical and technical support for medical diagnosis and treatment.

Optimization of Laser-Based Method to Conduct Skin Ablation in Zebrafish and Development of Deep Learning-Based Method for Skin Wound-Size Measurement

Skin plays an important role as a defense mechanism against environmental pathogens in organisms such as humans or animals. Once the skin integrity is disturbed by a wound, pathogens can penetrate easily into a deeper part of the body to induce disease. By this means, it is important for the skin to regenerate quickly upon injury to regain its protective barrier function. Traditionally, scientists use rodents or mammals as experimental animals to study skin wound healing. However, due to concerns about animal welfare and increasing costs of laboratory animals, such as rodents, scientists have considered alternative methods of implementing replace, reduce, and refine (3Rs) in experimentation. Moreover, several previous studies on skin wound healing in fish used relatively expensive medical-grade lasers with a low calculation efficiency of the wound area, which led to human judgment errors. Thus, this study aimed to develop a new alternative model for skin wound healing by utilizing zebrafish together with a new rapid and efficient method as an alternative in investigating skin wound healing. First, in order to fulfill the 3Rs concept, the pain in the tested zebrafish was evaluated by using a 3D locomotion assay. Afterward, the obtained behavior data were analyzed using the Kruskal–Wallis test, followed by Dunn’s multiple comparisons tests; later, 3 watts was chosen as the power for the laser, since the wound caused by the laser at this power did not significantly alter zebrafish swimming behaviors. Furthermore, we also optimized the experimental conditions of zebrafish skin wound healing using a laser engraving machine, which can create skin wounds with a high reproducibility in size and depth. The wound closure of the tested zebrafish was then analyzed by using a two-way ANOVA, and presented in 25%, 50%, and 75% of wound-closure percentages. After imparting wounds to the skin of the zebrafish, wound images were collected and used for deep-learning training by convolutional neural networks (CNNs), either the Mask-RCNN or U-Net, so that the computer could calculate the area of the skin wounds in an automatic manner. Using ImageJ manual counting as a gold standard, we found that the U-Net performance was better than the Mask RCNN for zebrafish skin wound judgment. For proof-of-concept validation, a U-Net trained model was applied to study and determine the effect of different temperatures and the administration of antioxidants on the skin wound-healing kinetics. Results showed a significant positive correlation between the speed of wound closure and the exposure to different temperatures and administration of antioxidants. Taken together, the laser-based skin ablation and deep learning-based wound-size measurement methods reported in this study provide a faster, reliable, and reduced suffering protocol to conduct skin wound healing in zebrafish for the first time.

Triplet attention fusion module: A concise and efficient channel attention module for medical image segmentation

Concise segmentation of medical images is vital for diagnosing and treating diseases. In recent years, convolutional neural networks (CNNs) have yielded satisfactory results in the field of medical image segmentation, and researchers are striving to improve the segmentation performance of CNNs by using the channel attention mechanism. However, most existing channel attention modules have numerous parameters and use a single compression and activation function in the process of attention realization. This limits the effect of attention promotion. To solve this problem, in this paper, we propose the triplet attention fusion (TAF) module. It combines direct and indirect mapping to achieve effective fusion of global-local information. In addition, this method has low computational complexity. Because of its lightweight and convenient advantages, the TAF module can be integrated into the existing semantic segmentation networks. The experimental results on ISIC-2018 and LiTS datasets show that TAF can greatly improve the performance of the network. Compared with other attention modules, it introduces fewer parameters. In addition, the performance of UNet with TAF module is close to or even surpasses that of the other state-of-the-art network models. It proves the superiority and effectiveness of our proposed module in medical image segmentation.

Mixed-Sized Biomedical Image Segmentation Based on U-Net Architectures

Convolutional neural networks (CNNs) are becoming increasingly popular in medical Image Segmentation. Among them, U-Net is a widely used model that can lead to cutting-edge results for 2D biomedical Image Segmentation. However, U-Net performance can be influenced by many factors, such as the size of the training dataset, the performance metrics used, the quality of the images and, in particular, the shape and size of the organ to be segmented. This could entail a loss of robustness of the U-Net-based models. In this paper, the performance of the considered networks is determined by using the publicly available images from the 3D-IRCADb-01 dataset. Different organs with different features are considered. Experimental results show that the U-Net-based segmentation performance decreases when organs with sparse binary masks are considered. The solution proposed in this paper, based on automated zooming of the parts of interest, allows improving the performance of the segmentation model by up to 20% in terms of Dice coefficient metric, when very sparse segmentation images are used, without affecting the cost of the learning process.

A deep learning classification approach using high spatial satellite images for detection of built-up areas in rural zones: Case study of Souss-Massa region - Morocco

The buildings in the rural areas of Morocco exist in various shapes and sizes. They are randomly distributed and are generally constructed of primary materials such as clay, wood, and tin. For these reasons, their detection is generally difficult and inaccurate with optical satellite imagery and traditional image processing techniques, particularly in rural settlements. New approaches, particularly those of Deep Learning, are called for testing their contribution to the detection of buildings and settlements in rural areas. This study aims to detect and map the settlements in rural areas in the Souss-Massa region using Sentinel-2 satellite images, based on deep Learning algorithms. First, we tested the result of the convolutional neural network architecture UNet. Then, to evaluate the impact of filters number on the performance of UNet, we increased the number of filters in the convolution layer. And third, we implement the deep Residual UNet (ResUNet). To evaluate the quality of tested models, special metrics, such as accuracy, precision, recall, F1-score, and the ROC curve are used. The obtained precision of 87% of precision and 54% of F1-score for the UNet with an increased number of filters outperforms the other algorithms UNet and ResUNet, which have 86.2% of precision and 81.2% precision, respectively. We compared our perception to those of other related studies conducted to extract buildings or settlements in rural areas using high to very high-resolution images and machine learning and deep learning algorithms. Our results show that the performance of settlement detection in rural areas using deep learning is affected by the quality and quantity of the model training base images and the number of filters used in the convolution layer.

DPCFN: Dual path cross fusion network for medical image segmentation

Thanks to the better performance of U-Net in medical segmentation, many U-Net variants have emerged one after another, but U-Net has the non-negligible drawback that it cannot accurately segment low-level features such as edge regions of images, and in addition this simple skip connection of U-Net itself is still a challenge for global information modeling. To solve the above problems, Channel-wise Cross Fusion Transformer (CCT) and Channel-wise Cross Attention (CCA) are introduced on the basis of U-Net, where CCT is used for cross fusion of U-Net encoders and Transformer, and CCA interacts the fused features with the decoder features to eliminate semantic gaps, naming the network Trans-Net. Another branch network SeU-Net is built to capture details and edge regions, and SE-Attention is added at the skip joints of the network to reinforce important features. The two branches interact through a Cross Residual Feature Block (CRFB). By testing on five datasets, it was experimentally demonstrated that the method proposed in this paper has more accurate segmentation performance.

Accelerating 2D and 3D frequency-domain seismic wave modeling through interpolating frequency-domain wavefields by deep learning

An attractive feature of finite-difference modeling in the frequency domain is the low recomputation cost to simulate seismic waves for many sources through the same velocity model. However, it is time consuming if many frequencies are involved for solving the linear wave equations, particularly, for large 3D velocity models. We propose accelerating the modeling by applying deep learning. Because similarity appears among the wavefields in near frequencies, we can extract similar features and reconstruct the unknown wavefields by deep learning. We compute fewer frequency-domain wavefields with a large frequency interval by conventional modeling methods, such as finite-difference methods, and then interpolate more frequency-domain wavefields with a small frequency interval by applying deep learning. Numerical examples demonstrate that the U-Net, which is trained by the data on 10 2D simple-layered models, can be used to interpolate the data on SEG Advanced Modeling models and also can be used to interpolate the data on the 3D layered model and 3D overthrust model. A series of tests on perturbation models prove that the U-Net performance is still good when the root-mean-square value of the velocity model perturbation relative to the initial model is up to 15%. Compared with traditional modeling methods, the computational time of interpolating by applying deep learning is negligible, especially for 3D models. After the network is trained, the acceleration strategy helps to reduce the runtime by approximately 50% for a modeling problem because the U-Net can be used to interpolate 50% of the data that would otherwise need to be modeled. Although training the model will take some time, good generalization of the U-Net, especially for 3D problems, enhances its benefits on the application for frequency-domain forward modeling.

Deep Learning-Based Multi-Label Tissue Segmentation and Density Assessment from Mammograms

Objectives: Breast cancer is the most commonly diagnosed type of cancer among women and a common cause of cancer-related deaths. Early diagnosis and treatment of breast cancer is critical in disease prognosis. Breast density is known to have a correlation with breast cancer. In recent years, there has been an increasing interest in the investigation of computer-aided methods for early diagnosis of breast cancer. In this study, a new fully-automated deep learning-based cascaded model was proposed for breast density assessment. In the first stage, the segmentation of adipose, fibroglandular, and pectoral muscle tissues from the digitized film mammograms of the Digital Database for Screening Mammography (DDSM) was investigated using various types of U-nets. Features extracted from the breast tissue segmentation predictions were then used to assess breast density in the second stage. Material and methods: 66 and 296 mediolateral oblique mammograms were selected from DDSM dataset for segmentation and breast density assessment systems, respectively. Different U-nets with varying number of layers and filters were implemented and the model having the highest performance was determined. U-net performance was investigated using categorical cross-entropy, Dice, Tversky, Focal Tversky, and logarithmic cosine-hyperbolic Dice loss functions. The performances of U-nets having different types of connections were investigated. The performances of U-nets having pre-trained weights from VGG16, VGG19, and ResNet50 networks in the encoding path were also investigated. Segmentation results were improved by using an image processing pipeline based on morphological operators. Segmentation performance was presented in terms of accuracy, balanced accuracy, intersection over union, and Dice's similarity coefficient (DSC) metrics. The segmentation system predictions were then used to estimate mammographic density using a machine learning pipeline by extracting features related to the fibroglandular tissue percentage. Results: Using ResNet50-U-net on the test data, average DSC scores of 82.71%, 73.39%, and 95.30% were obtained for adipose, fibroglandular, and pectoral muscle tissue segmentation, respectively. The mammogram segmentation results are 3%-12% better than the current state-of-the-art DSC in the literature when considering all of the foreground tissues concurrently. A breast density classification accuracy of 76.01% was achieved on a separate mammogram dataset, which is comparable to the recent studies in the literature. Conclusion: The proposed system can be used for automatic segmentation of mammogram into adipose, fibroglandular, and pectoral muscle tissues. The segmentation model enables the estimation of the fibroglandular-adipose tissue interface, which is recently found to be an important region for breast cancer investigations. The proposed fully-automatic breast density assessment system has a comparable performance to the ones in the literature.

Evaluation of a deep-learning model for multispectral remote sensing of land use and crop classification

High-resolution deep-learning-based remote-sensing imagery analysis has been widely used in land-use and crop-classification mapping. However, the influence of composite feature bands, including complex feature indices arising from different sensors on the backbone, patch size, and predictions in transferable deep models require further testing. The experiments were conducted in six sites in Henan province from 2019 to 2021. This study sought to enable the transfer of classification models across regions and years for Sentinel-2A (10-m resolution) and Gaofen PMS (2-m resolution) imagery. With feature selection and up-sampling of small samples, the performance of UNet++ architecture on five backbones and four patch sizes was examined. Joint loss, mean Intersection over Union (mIoU), and epoch time were analyzed, and the optimal backbone and patch size for both sensors were Timm-RegNetY-320 and 256 × 256, respectively. The overall accuracy and F1 scores of the Sentinel-2A predictions ranged from 96.86% to 97.72% and 71.29% to 80.75%, respectively, compared to 75.34%–97.72% and 54.89%–73.25% for the Gaofen predictions. The accuracies of each site indicated that patch size exerted a greater influence on model performance than the backbone. The feature-selection-based predictions with UNet++ architecture and up-sampling of minor classes demonstrated the capabilities of deep-learning generalization for classifying complex ground objects, offering improved performance compared to the UNet, Deeplab V3+, Random Forest, and Object-Oriented Classification models. In addition to the overall accuracy, confusion matrices, precision, recall, and F1 scores should be evaluated for minor land-cover types. This study contributes to large-scale, dynamic, and near-real-time land-use and crop mapping by integrating deep learning and multi-source remote-sensing imagery

Performance evaluation of segmentation methods for assessing the lens of the frog Thoropa miliaris from synchrotron-based phase-contrast micro-CT images.

In the context of synchrotron microtomography using propagation-based phase-contrast imaging (XSPCT), we evaluated the performance of semiautomatic and automatic image segmentation of soft biological structures by means of Dice Similarity Coefficient (DSC) and volume quantification. We took advantage of the phase-contrast effects of XSPCT to provide enhanced object boundaries and improved visualization of the lenses of the frog Thoropa miliaris. Then, we applied semiautomatic segmentation methods 1 and 2 (Interpolation and Watershed, respectively) and method 3, an automatic segmentation algorithm using the U-Net architecture, to the reconstructed images. DSC and volume quantification of the lenses were used to quantify the performance of image segmentation methods. Comparing the lenses segmented by the three methods, the most pronounced difference in volume quantification was between methods 1 and 3: a reduction of 4.24%. Method 1, 2 and 3 obtained the global average DSC of 97.02%, 95.41% and 89.29%, respectively. Although it obtained the lowest DSC, method 3 performed the segmentation in a matter of seconds, while the semiautomatic methods had the average time to segment the lenses around 1h and 30min. Our results suggest that the performance of U-Net was impaired due to the irregularities of the ROI edges mainly in its lower and upper regions, but it still showed high accuracy (DSC=89.29%) with significantly reduced segmentation time compared to the semiautomatic methods. Besides, with the present work we have established a baseline for future assessments of Deep Neural Networks applied to XSPCT volumes.

Learning U-Net Based Multi-Scale Features in Encoding-Decoding for MR Image Brain Tissue Segmentation.

Accurate brain tissue segmentation of MRI is vital to diagnosis aiding, treatment planning, and neurologic condition monitoring. As an excellent convolutional neural network (CNN), U-Net is widely used in MR image segmentation as it usually generates high-precision features. However, the performance of U-Net is considerably restricted due to the variable shapes of the segmented targets in MRI and the information loss of down-sampling and up-sampling operations. Therefore, we propose a novel network by introducing spatial and channel dimensions-based multi-scale feature information extractors into its encoding-decoding framework, which is helpful in extracting rich multi-scale features while highlighting the details of higher-level features in the encoding part, and recovering the corresponding localization to a higher resolution layer in the decoding part. Concretely, we propose two information extractors, multi-branch pooling, called MP, in the encoding part, and multi-branch dense prediction, called MDP, in the decoding part, to extract multi-scale features. Additionally, we designed a new multi-branch output structure with MDP in the decoding part to form more accurate edge-preserving predicting maps by integrating the dense adjacent prediction features at different scales. Finally, the proposed method is tested on datasets MRbrainS13, IBSR18, and ISeg2017. We find that the proposed network performs higher accuracy in segmenting MRI brain tissues and it is better than the leading method of 2018 at the segmentation of GM and CSF. Therefore, it can be a useful tool for diagnostic applications, such as brain MRI segmentation and diagnosing.

Magnetic resonance image (MRI) synthesis from brain computed tomography (CT) images based on deep learning methods for magnetic resonance (MR)-guided radiotherapy.

Precise patient setup is critical in radiation therapy. Medical imaging plays an essential role in patient setup. As compared to computed tomography (CT) images, magnetic resonance image (MRI) has high contrast for soft tissues, which becomes a promising imaging modality during treatment. In this paper, we proposed a method to synthesize brain MRI images from corresponding planning CT (pCT) images. The synthetic MRI (sMRI) images can be used to align with positioning MRI (pMRI) equipped by an MRI-guided accelerator to account for the disadvantages of multi-modality image registration. Several deep learning network models were applied to implement this brain MRI synthesis task, including CycleGAN, Pix2Pix model, and U-Net. We evaluated these methods using several metrics, including mean absolute error (MAE), mean squared error (MSE), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR). In our experiments, U-Net with L1+L2 loss achieved the best results with the lowest overall average MAE of 74.19 and MSE of 1.035*104, respectively, and produced the highest SSIM of 0.9440 and PSNR of 32.44. Quantitative comparisons suggest that the performance of U-Net, a supervised deep learning method, is better than the performance of CycleGAN, a typical unsupervised method, in our brain MRI synthesis procedure. The proposed method can convert pCT/pMRI multi-modality registration into mono-modality registration, which can be used to reduce registration error and achieve a more accurate patient setup.

Surgical Navigation System for Transsphenoidal Pituitary Surgery Applying U-Net-Based Automatic Segmentation and Bendable Devices

Conventional navigation systems used in transsphenoidal pituitary surgery have limitations that may lead to organ damage, including long image registration time, absence of alarms when approaching vital organs and lack of 3-D model information. To resolve the problems of conventional navigation systems, this study proposes a U-Net-based, automatic segmentation algorithm for optical nerves and internal carotid arteries, by training patient computed tomography angiography images. The authors have also developed a bendable endoscope and surgical tool to eliminate blind regions that occur when using straight, rigid, conventional endoscopes and surgical tools during transsphenoidal pituitary surgery. In this study, the effectiveness of a U-Net-based navigation system integrated with bendable surgical tools and a bendable endoscope has been demonstrated through phantom-based experiments. In order to measure the U-net performance, the Jaccard similarity, recall and precision were calculated. In addition, the fiducial and target registration errors of the navigation system and the accuracy of the alarm warning functions were measured in the phantom-based environment.