Deep U-Net Research Articles

Assessment of Self-Supervised Denoising Methods for Esophageal Speech Enhancement

Esophageal speech (ES) is a pathological voice that is often difficult to understand. Moreover, acquiring recordings of a patient’s voice before a laryngectomy proves challenging, thereby complicating enhancing this kind of voice. That is why most supervised methods used to enhance ES are based on voice conversion, which uses healthy speaker targets, things that may not preserve the speaker’s identity. Otherwise, unsupervised methods for ES are mostly based on traditional filters, which cannot alone beat this kind of noise, making the denoising process difficult. Also, these methods are known for producing musical artifacts. To address these issues, a self-supervised method based on the Only-Noisy-Training (ONT) model was applied, consisting of denoising a signal without needing a clean target. Four experiments were conducted using Deep Complex UNET (DCUNET) and Deep Complex UNET with Complex Two-Stage Transformer Module (DCUNET-cTSTM) for assessment. Both of these models are based on the ONT approach. Also, for comparison purposes and to calculate the evaluation metrics, the pre-trained VoiceFixer model was used to restore the clean wave files of esophageal speech. Even with the fact that ONT-based methods work better with noisy wave files, the results have proven that ES can be denoised without the need for clean targets, and hence, the speaker’s identity is retained.

Evaluation of OCT biomarker changes in treatment-naive neovascular AMD using a deep semantic segmentation algorithm.

To determine real-life quantitative changes in OCT biomarkers in a large set of treatment naive patients in a real-life setting undergoing anti-VEGF therapy. For this purpose, we devised a novel deep learning based semantic segmentation algorithm providing the first benchmark results for automatic segmentation of 11 OCT features including biomarkers for neovascular age-related macular degeneration (nAMD). Training of a Deep U-net based semantic segmentation ensemble algorithm for state-of-the-art semantic segmentation performance which was used to analyze OCT features prior to, after 3 and 12 months of anti-VEGF therapy. High F1 scores of almost 1.0 for neurosensory retina and subretinal fluid on a separate hold-out test set with unseen patients. The algorithm performed worse for subretinal hyperreflective material and fibrovascular PED, on par with drusenoid PED, and better in segmenting fibrosis. In the evaluation of treatment naive OCT scans, significant changes occurred for intraretinal fluid (mean: 0.03 µm3 to 0.01 µm3, p < 0.001), subretinal fluid (0.08 µm3 to 0.01 µm3, p < 0.001), subretinal hyperreflective material (0.02 µm3 to 0.01 µm3, p < 0.001), fibrovascular PED (0.12 µm3 to 0.09 µm3, p = 0.02) and central retinal thickness C0 (225.78 µm3 to 169.40 µm3). The amounts of intraretinal fluid, fibrovascular PED, and ERM were predictive of poor outcome. The segmentation algorithm allows efficient volumetric analysis of OCT scans. Anti-VEGF provokes most potent changes in the first 3 months while a gradual loss of RPE hints at a progressing decline of visual acuity. Additional research is required to understand how these accurate OCT predictions can be leveraged for a personalized therapy regimen.

Individualised prediction of longitudinal change in multimodal brain imaging

Abstract It remains largely unknown whether individualised longitudinal changes of brain imaging features can be predicted based only on the baseline brain images. This would be of great value, for example, for longitudinal data imputation, longitudinal brain-behaviour associations, and early prediction of brain-related diseases. We explore this possibility using longitudinal data of multiple modalities from UK Biobank brain imaging, with around 3,500 subjects. As baseline and follow-up images are generally similar in the case of short follow-up time intervals (e.g., 2 years), a simple copy of the baseline image may have a very good prediction performance. Therefore, for the first time, we propose a new mathematical framework for guiding the longitudinal prediction of brain images, providing answers to fundamental questions: (1) what is a suitable definition of longitudinal change; (2) how to detect the existence of changes; (3) what is the “null” prediction performance; and (4) can we distinguish longitudinal change prediction from simple data denoising. Building on these, we designed a deep U-Net based model for predicting longitudinal changes in multimodal brain images. Our results show that the proposed model can predict to a modest degree individualised longitudinal changes in almost all modalities, and outperforms other potential models. Furthermore, compared with the true longitudinal changes computed from real data, the predicted longitudinal changes have a similar or even improved accuracy in predicting subjects’ non-imaging phenotypes, and have a high between-subject discriminability. Our study contributes a new theoretical framework for longitudinal brain imaging studies, and our results show the potential for longitudinal data imputation, along with highlighting several caveats when performing longitudinal data analysis.

Assessing and segmenting salt-affected soils using in-situ EC measurements, remote sensing, and a modified deep learning MU-NET convolutional neural network

A study revealed that the Siwa Oasis faces high soil salinity, which negatively impacts agricultural areas and crop productivity, despite its significant economic and agricultural importance. The current research proposes an approach to detect and segment salinity and vegetation areas at Siwa Oasis, Egypt, by combining remote sensing and building a deep learning neural network model-based U-NET algorithm to detect salinity change areas, anticipate further degradation, and predict soil quality indicators. To locate changes among the available images, standard image improvement, classification, and change detection methods have been used. We applied a deep learning modified U-Net (MU-NET) algorithm to segment and produce salinity maps. The MU-NET architecture is a two-level nested U-structure merged with a residual U-block (RUb), which consists of an encoder and a decoder. We applied RUb, which consists of several layers and skip connections. Different combinations of the salinity and vegetation indices were added to the original image to improve segmentation precision. The model was validated and trained using actual data samples collected over a 10-year period from the Landsat 8 satellite, which can monitor and analyse present land cover changes. The dataset consisted of 91 OLI and TIRS spectral images. Each one consists of eleven bands with a spatial resolution of 30 m for bands 1 to 7 and 9. A field survey was used as the main source of data for comparing the proposed model's outputs to assess the error rate. The study region is experiencing an increase in soil salinity in all directions, particularly with regard to the spatial distribution of saline soils, not just the quantitative increase in salt-affected soils. These findings supported the acceleration of soil salinization and vegetation death. The proposed model achieved the highest performance results among the other models and literature and was based on applying method 12 using 13 image layers, with the highest accuracies of 91.27% and 90.83% for salinity and vegetation, respectively.

A new optimization approach based on neural architecture search to enhance deep U-Net for efficient road segmentation

Neural Architecture Search (NAS) has significantly improved the accuracy of image classification and segmentation. However, these methods concentrate on finding segmentation structures for natural or medical applications. In this study, we introduce a NAS approach based on gradient optimization to identify ideal cell designs for road segmentation. To the best of our knowledge, this work represents the first application of gradient-based NAS to road extraction. Taking insight from the U-Net model and its successful variations in different image segmentation tasks, we propose NAS-enhanced U-Net, illustrated by an equal number of cells in both encoder and decoder levels. While cross-entropy combined with dice loss is commonly used in many segmentation tasks, road extraction brings up a unique challenge due to class imbalance. To address this, we introduce a combination of loss function. This function merges cross-entropy with weighted Dice loss, focusing on elevating the importance of the road class by assigning it a weight (⍵), while background Dice values are disregarded. The results indicate that the optimal weight for the proposed model equals 2. Additionally, our work challenges the assumption that increased model parameters or depth inherently leads to improved performance. Therefore, we establish search spaces 2,3,4,5,6,7 and 8 to automatically choose the optimal depth for model. We present promising segmentation results for our proposed method, achieved without any pretraining on the Massachusetts road dataset. Furthermore, these results are compared with those of 14 models categorized into four groups: U-Net, Segnet, FCN8, and Nas-U-Net.

High-Resolution Model for Segmenting and Predicting Brain Tumor Based on Deep UNet with Multi Attention Mechanism

High-Resolution Model for Segmenting and Predicting Brain Tumor Based on Deep UNet with Multi Attention Mechanism

Deep U-NET Based Heating Film Defect Inspection System

Deep U-NET Based Heating Film Defect Inspection System

A Progressive UNDML Framework Model for Breast Cancer Diagnosis and Classification

According to recent research, it is studied that the second most common cause of death for women worldwide is breast cancer. Since it can be incredibly difficult to determine the true cause of breast cancer, early diagnosis is crucial to lowering the disease's fatality rate. Early cancer detection raises the chance of survival by up to 8 %. Radiologists look for irregularities in breast images collected from mammograms, X-rays, or MRI scans. Radiologists of all levels struggle to identify features like lumps, masses, and micro-calcifications, which leads to high false-positive and false-negative rates. Recent developments in deep learning and image processing give rise to some optimism for the creation of improved applications for the early diagnosis of breast cancer. A methodological study was carried out in which a new Deep U-Net Segmentation based Convolutional Neural Network, named UNDML framework is developed for identifying and categorizing breast anomalies. This framework involves the operations of preprocessing, quality enhancement, feature extraction, segmentation, and classification. Preprocessing is carried out in this case to enhance the quality of the breast picture input. Consequently, the Deep U-net segmentation methodology is applied to accurately segment the breast image for improving the cancer detection rate. Finally, the CNN mechanism is utilized to categorize the class of breast cancer. To validate the performance of this method, an extensive simulation and comparative analysis have been performed in this work. The obtained results demonstrate that the UNDML mechanism outperforms the other models with increased tumor detection rate and accuracy

Performance Comparison ConvDeconvNet Algorithm Vs. UNET for Fish Object Detection

The precise identification and localization of fish entities within visual data is essential in diverse domains, such as marine biology and fisheries management, within computer vision. This study provides a thorough performance evaluation of two prominent deep learning algorithms, ConvDeconvNet and UNET, in the context of fish object detection. Both models are assessed using a dataset comprising a wide range of fish species, considering various factors, including accuracy of detection, speed of processing, and complexity of the model. The findings demonstrate that ConvDeconvNet exhibits superior performance in terms of detection accuracy, attaining a noteworthy degree of precision and recall in identifying fish entities. In contrast, the UNET model displays a notable advantage in terms of processing speed owing to its distinctive architectural design, rendering it a viable option for applications requiring real-time performance. The discourse surrounding the trade-off between accuracy and speed is examined, offering valuable perspectives for algorithm selection following specific criteria. Furthermore, this study highlights the significance of incorporating a diverse range of datasets for training and testing purposes when utilizing these models, as it significantly influences their overall performance. This study makes a valuable contribution to the continuous endeavors to improve the detection of fish objects in underwater images. It provides a thorough evaluation and comparison of ConvDeconvNet and UNET, thereby assisting researchers and practitioners in making well-informed decisions regarding selecting these models for their specific applications.

Fully Automated Construction of a Deep U-Net Network Model for Medical Image Segmentation

Fully Automated Construction of a Deep U-Net Network Model for Medical Image Segmentation

Advance brain tumor segmentation using feature fusion methods with deep U-Net model with CNN for MRI data

In modern healthcare, the precision of medical image segmentation holds immense significance for diagnosis and treatment planning. Deep learning techniques, such as CNNs, UNETs, and Transformers, have revolutionized this field by automating the previously labor-intensive manual segmentation processes. However, challenges like intricate structures and indistinct features persist, leading to accuracy issues. Researchers are diligently addressing these challenges to further unlock the potential of medical image segmentation in healthcare transformation. To enhance the precision of brain tumor MRI image segmentation, our study introduces three novel feature-enhanced hybrid UNet models (FE-HU-NET): FE1-HU-NET, FE2-HU-NET, and FE3-HU-NET. Our approach encompasses three main aspects. Initially, we emphasize feature enhancement during the image preprocessing stage. We apply distinct image enhancement techniques—CLAHE, MHE, and MBOBHE—to each model. Secondly, we tailor the architecture of the UNet model to enhance segmentation results, focusing on a personalized layered design. Lastly, we employ a CNN model in post-processing to refine segmentation outcomes through additional convolutional layers. The HU-Net module, shared across the three models, integrates a customized UNet layer and a CNN. We also introduce an alternative feature-enhanced variant, FE4-HU-NET, utilizing the DeepLABv3 model. Incorporating CLAHE for image enhancement and bolstered by CNN layers, this variant offers a distinct approach. Rigorous experimentation underscores the excellence of our proposed framework in distinguishing complex brain tissues, surpassing current state-of-the-art models. Impressively, we achieve accuracy rates exceeding 99% across two publicly available datasets. Performance metrics such as the Jaccard index, sensitivity, and specificity further substantiate the effectiveness of our Hybrid U-Net model.

Pattern recognition of grooves in human lips for improved authentication in cyber-physical systems using U-Net architecture

Latent Lip groove application is been a notable topic in forensic applications like crime and other investigations. The detection of lip movement is been a challenging task since it is a smaller integral part of the human face. The conventional models operate on the available public or private dataset but it is constrained to the large population and unconstrained environment. The study aims at developing a deep learning model in a multimodal system using the deep U-Net Convolutional Neural Network architecture. It also aims at improving biometric authentication through a deep pattern recognition that involves the feature extraction of grooves present in the human lips. An examination of grooves present in the input lip image is conducted by the present system to check the authenticity of the person entering the cyber-physical systems. The lip images are collected from the public security cameras via high-definition cameras in crowded areas that help the proposed method in forensic investigation and further, it considers various unconstrained scenarios to improve the efficacy of the system. The study involves initially pre-processing of lip image, and feature extraction of lip grooves to improve the efficacy of the lip trait. The simulation is conducted on the MATLAB tool to examine the efficacy of the model against various existing methods. Further, the study does not take into account the datasets available on the websites and lip images are only collected from a large set population in a real-time environment. The results of the simulation show that the proposed method achieves a higher degree of accuracy in extracting the grooves from the input lip images.

Physics-informed data-driven model for fluid flow in porous media

A physics-informed machine learning model is developed that can replace the numerical simulations of porous media. By learning the communications among grid cells in the numerical domain, this model is capable of accurately predicting flow fields for new sets of simulation runs. Because of the many possible random arrangements of particles and their orientation with respect to each other, generalization of permeability with high accuracy is not trivial — nor is it practical using conventional means. Furthermore, building a comprehensive database for different grain/pore arrangements is not possible because of the cost of running numerical simulations to generate the database that represent all possible arrangements. The objective is to predict grid-level flow fields in porous media as a priori to determining permeability of porous media. The rationale is that once the detailed grid-level dynamics can be accurately predicted using data-driven approach, for any configuration/topology of the porous media, the detailed dynamics could be predicted without any need for new expensive numerical simulation runs. In this work, a physics-informed deep learning model is developed by using the results from Lattice-Boltzmann simulations of randomly distributed circular grains to represent the porous media. A variety of porous structures are developed by changing the number, size, and location of circular grains. The deep U-Net and ResNet neural network architectures are combined to train a deep learning model which avoids the vanishing gradient issues. The continuity and momentum conservation equations are embedded into the loss function of deep learning architecture. Robustness of the developed model is then tested for numerous variations of porous media which have not been used for developing the model.

Our MapAI approach: focusing on data pipeline and loss functions

Precision building detection is a difficult challenge because resolution, lighting conditions, and image quality greatly influence the performance of machine learning models. Additionally, the building types, settlement structure, road structure, soil color and texture, vegetation, and car types can also affect image segmentation, making the solutions local or regional. In this paper, we describe the solution for the MapAI challenge submitted by the ATELIER team. We focused on two primary parts: data processing and loss functions. Our main insights were that the data can be effectively resampled by exploiting its structure and that the boundary intersection over union metric is very forgiving, not giving the network enough incentive to refine the borders. The segmentation was performed with a standard 5-level deep U-Net with an additional conditional random fields (CRF) denoiser.

Deep-Learning-Based Few-Angle Cardiac SPECT Reconstruction Using Transformer.

Convolutional neural networks (CNNs) have been extremely successful in various medical imaging tasks. However, because the size of the convolutional kernel used in a CNN is much smaller than the image size, CNN has a strong spatial inductive bias and lacks a global understanding of the input images. Vision Transformer, a recently emerged network structure in computer vision, can potentially overcome the limitations of CNNs for image-reconstruction tasks. In this work, we proposed a slice-by-slice Transformer network (SSTrans-3D) to reconstruct cardiac SPECT images from 3D few-angle data. To be specific, the network reconstructs the whole 3D volume using a slice-by-slice scheme. By doing so, SSTrans-3D alleviates the memory burden required by 3D reconstructions using Transformer. The network can still obtain a global understanding of the image volume with the Transformer attention blocks. Lastly, already reconstructed slices are used as the input to the network so that SSTrans-3D can potentially obtain more informative features from these slices. Validated on porcine, phantom, and human studies acquired using a GE dedicated cardiac SPECT scanner, the proposed method produced images with clearer heart cavity, higher cardiac defect contrast, and more accurate quantitative measurements on the testing data as compared with a deep U-net.

Edge U-Net: Brain tumor segmentation using MRI based on deep U-Net model with boundary information

Blood clots in the brain are frequently caused by brain tumors. Early detection of these clots has the potential to significantly lower morbidity and mortality in cases of brain cancer. It is thus indispensable for a proper brain tumor diagnosis and treatment that tumor tissue magnetic resonance images (MRI) be accurately segmented. Several deep learning approaches to the segmentation of brain tumor MRIs have been proposed, each having been designed to properly map out ‘boundaries’ and thus achieve highly accurate segmentation. This study introduces a deep convolution neural network (DCNN), named the Edge U-Net model, built as an encoder-decoder structure inspired by the U-Net architecture. The Edge U-Net model can more precisely localise tumors by merging boundary-related MRI data with the main data from brain MRIs. In the decoder phase, boundary-related information from original MRIs of different scales is integrated with the appropriate adjacent contextual information. A novel loss function was added to this segmentation model to improve performance. This loss function is enhanced with boundary information, allowing the learning process to produce more precise results. In the conducted experiments, a public dataset with 3064 T1-Weighted Contrast Enhancement (T1-CE) images of three well-known brain tumor types were used. The experiment demonstrated that the proposed framework achieved satisfactory Dice score values compared with state-of-art models, with highly accurate differentiation of brain tissues. It achieved Dice scores of 88.8 % for meningioma, 91.76 % for glioma, and 87.28 % for pituitary tumors. Computations of other performance metrics such as the Jaccard index, sensitivity, and specificity were also conducted. According to the results, the Edge U-Net model is a potential diagnostic tool that can be used to help radiologists more precisely segment brain tumor tissue images.

Towards Incompressible Laminar Flow Estimation Based on Interpolated Feature Generation and Deep Learning

For industrial design and the improvement of fluid flow simulations, computational fluid dynamics (CFD) solvers offer practical functions and conveniences. However, because iterative simulations demand lengthy computation times and a considerable amount of memory for sophisticated calculations, CFD solvers are not economically viable. Such limitations are overcome by CFD data-driven learning models based on neural networks, which lower the trade-off between accurate simulation performance and model complexity. Deep neural networks (DNNs) or convolutional neural networks (CNNs) are good illustrations of deep learning-based CFD models for fluid flow modeling. However, improving the accuracy of fluid flow reconstruction or estimation in these earlier methods is crucial. Based on interpolated feature data generation and a deep U-Net learning model, this work suggests a rapid laminar flow prediction model for inference of Naiver–Stokes solutions. The simulated dataset consists of 2D obstacles in various positions and orientations, including cylinders, triangles, rectangles, and pentagons. The accuracy of estimating velocities and pressure fields with minimal relative errors can be improved using this cutting-edge technique in training and testing procedures. Tasks involving CFD design and optimization should benefit from the experimental findings.

Semantic segmentation of soil salinity using in-situ EC measurements and deep learning based U-NET architecture

Soil salinity may occur naturally through pedogenetic processes or may result from abiotic factors resulting from human activities such as irrigation with poor quality water, lack of drainage and land development. The infertility of agricultural lands due to soil salinity brings many economic and social problems on a local and global scale. In this study, the salinity problem in the Harran Plain, one of the largest agricultural areas in Turkey, was investigated using deep learning-based U-NET algorithm. Different combinations of Normalized Difference Salinity Index (NDSI), Salinity Index I (SI), Salinity Index II (SII), and Normalized Difference Vegetation Index (NDVI) indices integrated into the RapidEye multispectral image to increase the segmentation accuracy. The most successful result (93.78% overall accuracy) was achieved when the algorithm was trained with 300 iterations and only the SII index was added to the original image. The same images were also segmented by the SVM method, and the U-NET deep learning architecture was able to detect soil salinity about 20–32% more accurately than SVM for three different 5-band test images. This study demonstrates that the delineation and mapping of soil salinity can be done automatically thanks to a deep learning (DL) network with greater accuracy, less time and effort, and without requiring continuous in-situ Electrical Conductivity (EC) measurements and repetitive supervised learning for each new satellite image.

Semantic segmentation of human cell nucleus using deep U-Net and other versions of U-Net models

ABSTRACT The deep learning models play an essential role in many areas, including medical image analysis. These models extract important features without human intervention. In this paper, we propose a deep convolution neural network, named as deep U-Net model, for the segmentation of the cell nucleus, a critical functional unit that determines the function and structure of the body. The nucleus contains all kinds of DNA, RNA, chromosomes, and genes governing all life activities, and its disorder may lead to different types of diseases such as cancer, heart disease, diabetes, Alzheimer’s, etc. If the nucleus structure is known correctly, diseases due to nucleus disorder may be detected early. It may also reduce the drug discovery time if the shape and size of the nucleus are known. We evaluate the performance of the proposed models on the nucleus segmentation data set used by the Data Science Bowl 2018 competition hosted by Kaggle. We compare its performance with that of the U-Net, Attention U-Net, R2U-Net, Attention R2U-Net, and both versions of the U-Net++ with and without supervision, in terms of loss, dice coefficient, dice loss, intersection over union, and accuracy. Our model performs better than the existing models.

A state-of-the-art technique to perform cloud-based semantic segmentation using deep learning 3D U-Net architecture

Glioma is the most aggressive and dangerous primary brain tumor with a survival time of less than 14 months. Segmentation of tumors is a necessary task in the image processing of the gliomas and is important for its timely diagnosis and starting a treatment. Using 3D U-net architecture to perform semantic segmentation on brain tumor dataset is at the core of deep learning. In this paper, we present a unique cloud-based 3D U-Net method to perform brain tumor segmentation using BRATS dataset. The system was effectively trained by using Adam optimization solver by utilizing multiple hyper parameters. We got an average dice score of 95% which makes our method the first cloud-based method to achieve maximum accuracy. The dice score is calculated by using Sørensen-Dice similarity coefficient. We also performed an extensive literature review of the brain tumor segmentation methods implemented in the last five years to get a state-of-the-art picture of well-known methodologies with a higher dice score. In comparison to the already implemented architectures, our method ranks on top in terms of accuracy in using a cloud-based 3D U-Net framework for glioma segmentation.