Variants Of U-Net Research Articles

A Multiscale Nonlocal Feature Extraction Network for Breast Lesion Segmentation in Ultrasound Images

Breast lesion segmentation in ultrasound images is of great importance since it can help us to characterize and localize lesion regions. However, low-quality imaging, blurred boundary, and variable lesion shapes bring challenges to accurate segmentation. In recent years, many U-Net variants have been proposed and successfully applied to breast lesion segmentation. However, these methods suffer from two limitations: (1) Ignoring the ability to capture rich global context information, and (2) Introducing extra complex operations. To alleviate these challenges, we propose a multiscale nonlocal feature extraction network (MNFE-Net) for accurately segmenting breast lesions. The core idea includes three points: (1) Parallel Encoder models long-range dependencies, (2) Multiscale Feature Module refines local features without introducing extra complex operations, and (3) Global Feature Guidance Module extracts global semantic information. MNFE-Net mainly has the following advantages: (1) The method has excellent performance for segmentation of malignant breast lesions, (2) The Parallel Encoder increases network parameters without significantly decreasing inference speed, and (3) The method is easy to understand and execute. Extensive experiment results with six state-of-the-art (SOTA) methods on three public breast ultrasound datasets demonstrate the superior segmentation performance of our proposed MNFE-Net.

Fractal, recurrent, and dense U-Net architectures with EfficientNet encoder for medical image segmentation.

U-Net is a deep learning technique that has made significant contributions to medical image segmentation. Although the accomplishments of deep learning algorithms in terms of image processing are evident, many challenges still need to be overcome to achieve human-like performance. One of the main challenges in building deeper U-Nets is black-box problems, such as vanishing gradients. Overcoming this problem allows us to develop neural networks with advanced network designs. We propose three U-Net variants, namely efficient R2U-Net, efficient dense U-Net, and efficient fractal U-Net, that can create highly accurate segmentation maps. The first part of our contribution makes use of EfficientNet to distribute resources in the network efficiently. The second part of our work applies the following layer connections to design the U-Net decoders: residual connections, dense connections, and fractal expansion. We apply EfficientNet as the encoder to our three decoders to design three conceivable models. The aforementioned three proposed deep learning models were tested on four benchmark datasets, including the CHASE DB1 and digital retinal images for vessel extraction (DRIVE) retinal image databases and the ISIC 2018 and HAM10000 dermoscopy image databases. We obtained the highest Dice coefficient of 0.8013, 0.8808, 0.8019, and 0.9295 for CHASE DB1, ISIC 2018, DRIVE, and HAM10000, respectively, and a Jaccard (JAC) score of 0.6686, 0.7870, 0.6694, and 0.8683 for CHASE DB1, ISIC 2018, DRIVE, and HAM10000, respectively. Statistical analysis revealed that the proposed deep learning models achieved better segmentation results compared with the state-of-the-art models. U-Net is quite an adaptable deep learning framework and can be integrated with other deep learning techniques. The use of recurrent feedback connections, dense convolution, residual skip connections, and fractal convolutional expansions allow for the design of improved deeper U-Net models. With the addition of EfficientNet, we can now leverage the performance of an optimally scaled classifier for U-Net encoders.

Residual UNet with Dual Attention—An ensemble residual UNet with dual attention for multi‐modal and multi‐class brain MRI segmentation

AbstractIn this paper, we have proposed a variant of UNet for brain magnetic resonance imaging (MRI) segmentation. The proposed model, termed as Residual UNet with Dual Attention (RUDA), addresses the two significant challenges of UNet: extraction of the complex features with unclear boundaries and the problem of over‐segmentation due to the redundancy caused by the skip connection usage. RUDA is constituted upon the residual blocks for extracting the complex structures. It Introduces attention into the skip connections to avoid redundancy and thereby the chance of over‐segmentation. Our model segments brain MRI into white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF) regions, which are considered crucial informative substructures for diagnosing neurological disorders such as Alzheimer's. It has been implemented in an ensemble manner to accommodate the multi‐sequence (T1‐weighted, IR, and T2‐FLAIR) scans. The empirical analysis shows that with an accuracy of 93.80%, RUDA outperforms the two baseline models: UNet (91.37%), ResUNet (91.44%).

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.

HMRNet: High and Multi-Resolution Network With Bidirectional Feature Calibration for Brain Structure Segmentation in Radiotherapy.

Accurate segmentation of Anatomical brain Barriers to Cancer spread (ABCs) plays an important role for automatic delineation of Clinical Target Volume (CTV) of brain tumors in radiotherapy. Despite that variants of U-Net are state-of-the-art segmentation models, they have limited performance when dealing with ABCs structures with various shapes and sizes, especially thin structures (e.g., the falx cerebri) that span only few slices. To deal with this problem, we propose a High and Multi-Resolution Network (HMRNet) that consists of a multi-scale feature learning branch and a high-resolution branch, which can maintain the high-resolution contextual information and extract more robust representations of anatomical structures with various scales. We further design a Bidirectional Feature Calibration (BFC) block to enable the two branches to generate spatial attention maps for mutual feature calibration. Considering the different sizes and positions of ABCs structures, our network was applied after a rough localization of each structure to obtain fine segmentation results. Experiments on the MICCAI 2020 ABCs challenge dataset showed that: 1) Our proposed two-stage segmentation strategy largely outperformed methods segmenting all the structures in just one stage; 2) The proposed HMRNet with two branches can maintain high-resolution representations and is effective to improve the performance on thin structures; 3) The proposed BFC block outperformed existing attention methods using monodirectional feature calibration. Our method won the second place of ABCs 2020 challenge and has a potential for more accurate and reasonable delineation of CTV of brain tumors.

SIP-UNet: Sequential Inputs Parallel UNet Architecture for Segmentation of Brain Tissues from Magnetic Resonance Images

Proper analysis of changes in brain structure can lead to a more accurate diagnosis of specific brain disorders. The accuracy of segmentation is crucial for quantifying changes in brain structure. In recent studies, UNet-based architectures have outperformed other deep learning architectures in biomedical image segmentation. However, improving segmentation accuracy is challenging due to the low resolution of medical images and insufficient data. In this study, we present a novel architecture that combines three parallel UNets using a residual network. This architecture improves upon the baseline methods in three ways. First, instead of using a single image as input, we use three consecutive images. This gives our model the freedom to learn from neighboring images as well. Additionally, the images are individually compressed and decompressed using three different UNets, which prevents the model from merging the features of the images. Finally, following the residual network architecture, the outputs of the UNets are combined in such a way that the features of the image corresponding to the output are enhanced by a skip connection. The proposed architecture performed better than using a single conventional UNet and other UNet variants.

Boosting Unsupervised Dorsal Hand Vein Segmentation with U-Net Variants

The identification of vascular network structures is one of the key fields of research in medical imaging. The segmentation of dorsal hand vein patterns form NIR images is not only the basis for reliable biometric identification, but would also provide a significant tool in assisting medical intervention. Precise vein extraction would help medical workers to exactly determine the needle entry point to efficiently gain intravenous access for different clinical purposes, such as intravenous therapy, parenteral nutrition, blood analysis and so on. It would also eliminate repeated attempts at needle pricks and even facilitate an automatic injection procedure in the near future. In this paper, we present a combination of unsupervised and supervised dorsal hand vein segmentation from near-infrared images in the NCUT database. This method is convenient due to the lack of expert annotations of publicly available vein image databases. The novelty of our work is the automatic extraction of the veins in two phases. First, a geometrical approach identifies tubular structures corresponding to veins in the image. This step is considered gross segmentation and provides labels (Label I) for the second CNN-based segmentation phase. We visually observe that different CNNs obtain better segmentation on the test set. This is the reason for building an ensemble segmentor based on majority voting by nine different network architectures (U-Net, U-Net++ and U-Net3+, all trained with BCE, Dice and focal losses). The segmentation result of the ensemble is considered the second label (Label II). In our opinion, the new Label II is a better annotation of the NCUT database than the Label I obtained in the first step. The efficiency of computer vision algorithms based on artificial intelligence algorithms is determined by the quality and quantity of the labeled data used. Furthermore, we prove this statement by training ResNet–UNet in the same manner with the two different label sets. In our experiments, the Dice scores, sensitivity and specificity with ResNet–UNet trained on Label II are superior to the same classifier trained on Label I. The measured Dice scores of ResNet–UNet on the test set increase from 90.65% to 95.11%. It is worth mentioning that this article is one of very few in the domain of dorsal hand vein segmentation; moreover, it presents a general pipeline that may be applied for different medical image segmentation purposes.

Half-UNet: A Simplified U-Net Architecture for Medical Image Segmentation

Medical image segmentation plays a vital role in computer-aided diagnosis procedures. Recently, U-Net is widely used in medical image segmentation. Many variants of U-Net have been proposed, which attempt to improve the network performance while keeping the U-shaped structure unchanged. However, this U-shaped structure is not necessarily optimal. In this article, the effects of different parts of the U-Net on the segmentation ability are experimentally analyzed. Then a more efficient architecture, Half-UNet, is proposed. The proposed architecture is essentially an encoder-decoder network based on the U-Net structure, in which both the encoder and decoder are simplified. The re-designed architecture takes advantage of the unification of channel numbers, full-scale feature fusion, and Ghost modules. We compared Half-UNet with U-Net and its variants across multiple medical image segmentation tasks: mammography segmentation, lung nodule segmentation in the CT images, and left ventricular MRI image segmentation. Experiments demonstrate that Half-UNet has similar segmentation accuracy compared U-Net and its variants, while the parameters and floating-point operations are reduced by 98.6 and 81.8%, respectively, compared with U-Net.

Automatic liver tumor segmentation used the cascade multi-scale attention architecture method based on 3D U-Net

Due to the complex structure of liver tumors and the low contrast with normal tissues make it still a challenging task to accurately segment liver tumors from CT images. To address these problems, we propose an end-to-end segmentation method for liver tumors. The method uses a cascade structure to improve the network's extraction of information. First, the Side-output Feature Fusion Attention block is used to fuse features at different levels and combine with attention mechanism to focus on important information. Then, the Atrous Spatial Pyramid Pooling Attention block is used to extract multi-scale semantic features. Finally, the Multi-scale Prediction Fusion block is used to fully fused the features captured at each layer of the network. To verify the performance of the proposed model and the effectiveness of each module, we evaluate it on LiTS and 3DIRCADb datasets and obtained Dice per Case of 0.665 and 0.719, respectively, and Dice Global of 0.812 and 0.784, respectively. The proposed method is compared with the basic model 3D U-Net, as well as some mainstream methods based on U-Net variants, and our method achieves better performance on the liver tumor segmentation task and is superior to most segmentation algorithms.

R2U++: a multiscale recurrent residual U-Net with dense skip connections for medical image segmentation

U-Net is a widely adopted neural network in the domain of medical image segmentation. Despite its quick embracement by the medical imaging community, its performance suffers on complicated datasets. The problem can be ascribed to its simple feature extracting blocks: encoder/decoder, and the semantic gap between encoder and decoder. Variants of U-Net (such as R2U-Net) have been proposed to address the problem of simple feature extracting blocks by making the network deeper, but it does not deal with the semantic gap problem. On the other hand, another variant UNET++ deals with the semantic gap problem by introducing dense skip connections but has simple feature extraction blocks. To overcome these issues, we propose a new U-Net based medical image segmentation architecture R2U++. In the proposed architecture, the adapted changes from vanilla U-Net are: (1) the plain convolutional backbone is replaced by a deeper recurrent residual convolution block. The increased field of view with these blocks aids in extracting crucial features for segmentation which is proven by improvement in the overall performance of the network. (2) The semantic gap between encoder and decoder is reduced by dense skip pathways. These pathways accumulate features coming from multiple scales and apply concatenation accordingly. The modified architecture has embedded multi-depth models, and an ensemble of outputs taken from varying depths improves the performance on foreground objects appearing at various scales in the images. The performance of R2U++ is evaluated on four distinct medical imaging modalities: electron microscopy, X-rays, fundus, and computed tomography. The average gain achieved in IoU score is 1.5 ± 0.37% and in dice score is 0.9 ± 0.33% over UNET++, whereas, 4.21 ± 2.72 in IoU and 3.47 ± 1.89 in dice score over R2U-Net across different medical imaging segmentation datasets.

Combining dense elements with attention mechanisms for 3D radiotherapy dose prediction on head and neck cancers.

PurposeExternal radiation therapy planning is a highly complex and tedious process as it involves treating large target volumes, prescribing several levels of doses, as well as avoiding irradiating critical structures such as organs at risk close to the tumor target. This requires highly trained dosimetrists and physicists to generate a personalized plan and adapt it as treatment evolves, thus affecting the overall tumor control and patient outcomes. Our aim is to achieve accurate dose predictions for head and neck (H&N) cancer patients on a challenging in‐house dataset that reflects realistic variability and to further compare and validate the method on a public dataset.MethodsWe propose a three‐dimensional (3D) deep neural network that combines a hierarchically dense architecture with an attention U‐net (HDA U‐net). We investigate a domain knowledge objective, incorporating a weighted mean squared error (MSE) with a dose‐volume histogram (DVH) loss function. The proposed HDA U‐net using the MSE‐DVH loss function is compared with two state‐of‐the‐art U‐net variants on two radiotherapy datasets of H&N cases. These include reference dose plans, computed tomography (CT) information, organs at risk (OARs), and planning target volume (PTV) delineations. All models were evaluated using coverage, homogeneity, and conformity metrics as well as mean dose error and DVH curves.ResultsOverall, the proposed architecture outperformed the comparative state‐of‐the‐art methods, reaching 0.95 (0.98) on D95 coverage, 1.06 (1.07) on the maximum dose value, 0.10 (0.08) on homogeneity, 0.53 (0.79) on conformity index, and attaining the lowest mean dose error on PTVs of 1.7% (1.4%) for the in‐house (public) dataset. The improvements are statistically significant (p<0.05) for the homogeneity and maximum dose value compared with the closest baseline. All models offer a near real‐time prediction, measured between 0.43 and 0.88 s per volume.ConclusionThe proposed method achieved similar performance on both realistic in‐house data and public data compared to the attention U‐net with a DVH loss, and outperformed other methods such as HD U‐net and HDA U‐net with standard MSE losses. The use of the DVH objective for training showed consistent improvements to the baselines on most metrics, supporting its added benefit in H&N cancer cases. The quick prediction time of the proposed method allows for real‐time applications, providing physicians a method to generate an objective end goal for the dosimetrist to use as reference for planning. This could considerably reduce the number of iterations between the two expert physicians thus reducing the overall treatment planning time.

Artificial intelligence-based technology to make a three-dimensional pelvic model for preoperative simulation of rectal cancer surgery using MRI.

A new technique that allows visualization of whole pelvic organs with high accuracy and usability is needed for preoperative simulation in advanced rectal cancer surgery. In this study, we developed an automated algorithm to create a three-dimensional (3D) model from pelvic MRI using artificial intelligence (AI) technology. This study included a total of 143 patients who underwent 3D MRI in a preoperative examination for rectal cancer. The training dataset included 133 patients, in which ground truth labels were created for pelvic vessels, nerves, and bone. A 3D variant of U-net was used for the network architecture. Ten patients who underwent lateral lymph node dissection were used as a validation dataset. The correctness of the vascular labelling was assessed for pelvic vessels and the Dice similarity coefficients calculated for pelvic bone. An automatic segmentation algorithm that extracts the artery, vein, nerve, and pelvic bone was developed, automatically producing a 3D image of the entire pelvis. The total time needed for segmentation was 133seconds. The success rate of the AI-based segmentation was 100% for the common and external iliac vessels, but the rates for the vesical vein (75%), superior gluteal vein (60%), or accessory obturator vein (63%) were suboptimal. Regarding pelvic bone, the average Dice similarity coefficient between manual and automatic segmentation was 0.97 (standard deviation 0.0043). Though there is room to improve the segmentation accuracy, the algorithm developed in this study can be utilized for surgical simulation in the treatment of advanced rectal cancer.

Super U-Net: A modularized generalizable architecture

ObjectiveTo develop and validate a novel convolutional neural network (CNN) termed “Super U-Net” for medical image segmentation. MethodsSuper U-Net integrates a dynamic receptive field module and a fusion upsampling module into the classical U-Net architecture. The model was developed and tested to segment retinal vessels, gastrointestinal (GI) polyps, skin lesions on several image types (i.e., fundus images, endoscopic images, dermoscopic images). We also trained and tested the traditional U-Net architecture, seven U-Net variants, and two non-U-Net segmentation architectures. K-fold cross-validation was used to evaluate performance. The performance metrics included Dice similarity coefficient (DSC), accuracy, positive predictive value (PPV), and sensitivity. ResultsSuper U-Net achieved average DSCs of 0.808±0.0210, 0.752±0.019, 0.804±0.239, and 0.877±0.135 for segmenting retinal vessels, pediatric retinal vessels, GI polyps, and skin lesions, respectively. The Super U-net consistently outperformed U-Net, seven U-Net variants, and two non-U-Net segmentation architectures (p < 0.05). ConclusionDynamic receptive fields and fusion upsampling can significantly improve image segmentation performance.

2K-Fold-Net and feature enhanced 4-Fold-Net for medical image segmentation

For segmenting medical images, U-Net has become a popular and effective tool. However, it also has some shortcomings in segmenting fuzzy boundaries and eliminating interferences. Improvements of the original U-Net have been proposed by many authors, resulting in many variants such as MultiResUNet, DoubleU-Net and W-Net. Based on the common characteristics of these structures, we propose in this work a generalized structure by multiplying the folds of a fully convolutional network (FCN) for even more times, and thus name it as “2K-Fold-Net”. The more folds in this structure provide more freedoms to create cross links between the neighboring folds. The influence of the fold-pair number K on its performance is also studied. The realizations with K up to 6 are compared to three other variants of cascaded U-Nets using the CVC-ClinicDB dataset. Then the special case “4-Fold-Net” is further empowered with the feature enhancing functionalities recently seen in the attention-aware feature enhancement method. This new net is hence named as “Enhanced-Feature-4-Fold-Net”, abbreviated as “EF3-Net”. Finally, 2K-Fold-Net and EF3-Net have been compared with U-Net, SegNet, DoubleU-Net, MultiResUNet and its variants using four challenging medical image datasets. The results have demonstrated that the proposed nets outperform the other variants of U-Net, even with slightly lower amount of parameters. The code is available on: https://github.com/raik7/EF3-Net.

Modality specific U-Net variants for biomedical image segmentation: a survey.

With the advent of advancements in deep learning approaches, such as deep convolution neural network, residual neural network, adversarial network; U-Net architectures are most widely utilized in biomedical image segmentation to address the automation in identification and detection of the target regions or sub-regions. In recent studies, U-Net based approaches have illustrated state-of-the-art performance in different applications for the development of computer-aided diagnosis systems for early diagnosis and treatment of diseases such as brain tumor, lung cancer, alzheimer, breast cancer, etc., using various modalities. This article contributes in presenting the success of these approaches by describing the U-Net framework, followed by the comprehensive analysis of the U-Net variants by performing (1) inter-modality, and (2) intra-modality categorization to establish better insights into the associated challenges and solutions. Besides, this article also highlights the contribution of U-Net based frameworks in the ongoing pandemic, severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) also known as COVID-19. Finally, the strengths and similarities of these U-Net variants are analysed along with the challenges involved in biomedical image segmentation to uncover promising future research directions in this area.

İnsan Embriyo Segmentasyonu için U-Net Tabanlı Modellerin Karşılaştırılması

The quality of human embryos produced during in vitro fertilization is conventionally graded by clinical embryologists and this process is time-consuming and prone to human error. Artificial intelligence methods may be used to grade images captured by time-lapse microscopy (TLM). Segmentation of embryos from the background of TLM images is an essential step for embryo quality assessment as the background of the embryo has various artifacts which may mislead the grading algorithms. In this study, we performed a comparative analysis of automated day-5 human embryo (blastocyst) image segmentation methods based on deep learning. Four fully convolutional deep models, including U-Net and its three variants, were created using the combination of two gradient descent-based optimizers and two-loss functions and compared to our proposed model. The experimental results on the test set confirmed that our customized Dilated Inception U-Net model with Adam optimizer and Dice loss outperformed other U-Net variants with Dice coefficient, Jaccard index, accuracy, and precision of 98.68%, 97.52%, 99.20%, and 98.52%, respectively.

Automated 3D U‐net based segmentation of neonatal cerebral ventricles from 3D ultrasound images

Intraventricular hemorrhaging (IVH) within cerebral lateral ventricles affects 20-30% of very low birth weight infants (<1500g). As the ventricles increase in size, the intracranial pressure increases, leading to post-hemorrhagic ventricle dilatation (PHVD), an abnormal enlargement of the head. The most widely used imaging tool for measuring IVH and PHVD is cranial two-dimensional (2D) ultrasound (US). Estimating volumetric changes over time with 2D US is unreliable due to high user variability when locating the same anatomical location at different scanning sessions. Compared to 2D US, three-dimensional (3D) US is more sensitive to volumetric changes in the ventricles and does not suffer from variability in slice acquisition. However, 3D US images require segmentation of the ventricular surface, which is tedious and time-consuming when done manually. A fast, automated ventricle segmentation method for 3D US would provide quantitative information in a timely manner when monitoring IVH and PHVD in pre-term neonates. To this end, we developed a fast and fully automated segmentation method to segment neonatal cerebral lateral ventricles from 3D US images using deep learning. Our method consists of a 3D U-Net ensemble model composed of three U-Net variants, each highlighting various aspects of the segmentation task such as the shape and boundary of the ventricles. The ensemble is made of a U-Net++, attention U-Net, and U-Net with a deep learning-based shape prior combined using a mean voting strategy. We used a dataset consisting of 190 3D US images, which was separated into two subsets, one set of 87 images contained both ventricles, and one set of 103 images contained only one ventricle (caused by limited field-of-view during acquisition). We conducted fivefold cross-validation to evaluate the performance of the models on a larger amount of test data; 165 test images of which 75 have two ventricles (two-ventricle images) and 90 have one ventricle (one-ventricle images). We compared these results to each stand-alone model and to previous works including, 2D multiplane U-Net and 2D SegNet models. Using fivefold cross-validation, the ensemble method reported a Dice similarity coefficient (DSC) of 0.720 ± 0.074, absolute volumetric difference (VD) of 3.7 ± 4.1 cm3 , and a mean absolute surface distance (MAD) of 1.14 ± 0.41mm on 75 two-ventricle test images. Using 90 test images with a single ventricle, the model after cross-validation reported DSC, VD, and MAD values of 0.806 ± 0.111, 3.5 ± 2.9 cm3 , and 1.37 ± 1.70mm, respectively. Compared to alternatives, the proposed ensemble yielded a higher accuracy in segmentation on both test data sets. Our method required approximately 5 s to segment one image and was substantially faster than the state-of-the-art conventional methods. Compared to the state-of-the-art non-deep learning methods, our method based on deep learning was more efficient in segmenting neonatal cerebral lateral ventricles from 3D US images with comparable or better DSC, VD, and MAD performance. Our dataset was the largest to date (190 images) for this segmentation problem and the first to segment images that show only one lateral cerebral ventricle.

Understanding the Role of Receptive Field of Convolutional Neural Network for Cloud Detection in Landsat 8 OLI Imagery

Deep semantic segmentation networks perform better in cloud detection of satellite imagery than traditional methods due to their ability to extract high-level features over a large receptive field. However, a large receptive field often leads to loss of spatial details and blurring of boundaries. Therefore, it is crucial to understand the role of the receptive field on the segmentation results, which has rarely been investigated for cloud detection tasks. This study, for the first time, explored the relationship between the receptive field size and the performance of a cloud detection network. Six typical networks commonly used for cloud detection and nine modified UNet variants with different depths, dilated convolutions, and skip connections were evaluated based on the Landsat 8 Biome (L8 Biome) dataset. The theoretical receptive field (TRF) and the effective receptive field (ERF) were introduced to measure the receptive field sizes of different networks. The results revealed a negative correlation between the ERF size and cloud segmentation accuracies for different cloud distributions and a relatively weak negative correlation between the TRF size and segmentation accuracies. Furthermore, ERFs were considerably smaller than the corresponding TRFs for most networks, implying that large-scale contextual information was not learned after training. This result indicates the importance of using networks with a small receptive field for cloud detection of Landsat 8 OLI imagery. Moreover, as the boundary accuracies are significantly lower than the region accuracies, future efforts should be devoted to addressing inaccurate boundary localization rather than exploring the contextual information over a large receptive field.

Detecting Small Anatomical Structures in 3D Knee MRI Segmentation by Fully Convolutional Networks

Accurately identifying the pixels of small organs or lesions from magnetic resonance imaging (MRI) has a critical impact on clinical diagnosis. U-net is the most well-known and commonly used neural network for image segmentation. However, the small anatomical structures in medical images cannot be well recognised by U-net. This paper explores the performance of the U-net architectures in knee MRI segmentation to find a relative structure that can obtain high accuracies for both small and large anatomical structures. To maximise the utilities of U-net architecture, we apply three types of components, residual blocks, squeeze-and-excitation (SE) blocks, and dense blocks, to construct four variants of U-net, namely U-net variants. Among these variants, our experiments show that SE blocks can improve the segmentation accuracies of small labels. We adopt DeepLabv3plus architecture for 3D medical image segmentation by equipping SE blocks based on this discovery. The experimental results show that U-net with SE block achieves higher accuracy in parts of small anatomical structures. In contrast, DeepLabv3plus with SE block performs better on the average dice coefficient of small and large labels.

SEGMENTATION OF ADDITIVE MANUFACTURING DEFECTS USING U-NET

Abstract Additive manufacturing (AM) provides design flexibility and allows rapid fabrications of parts with complex geometries. The presence of internal defects, however, can lead to deficit performance of the fabricated part. X-ray Computed Tomography (XCT) is a non-destructive inspection technique often used for AM parts. Although defects within AM specimens can be identified and segmented by manually thresholding the XCT images, the process can be tedious and inefficient, and the segmentation results can be ambiguous. The variation in the shapes and appearances of defects also poses difficulty in accurately segmenting defects. This paper describes an automatic defect segmentation method using U-Net based deep convolutional neural network (CNN) architectures. Several models of U-Net variants are trained and validated on an AM XCT image dataset containing pores and cracks, achieving a best mean intersection over union (IOU) value of 0.993. Performance of various U-Net models is compared and analyzed. Specific to AM porosity segmentation with XCT images, several techniques in data augmentation and model development are introduced. This work demonstrates that U-Net can be effectively applied for automatic segmentation of AM porosity from XCT images with high accuracy. The method can potentially help improve quality control of AM parts in an industry setting.