Advertisement

Application of an Improved U2-Net Model in Ultrasound Median Neural Image Segmentation

Open AccessPublished:September 24, 2022DOI:https://doi.org/10.1016/j.ultrasmedbio.2022.08.003

      Abstract

      To investigate whether an improved U2-Net model could be used to segment the median nerve and improve segmentation performance, we performed a retrospective study with 402 nerve images from patients who visited Huashan Hospital from October 2018 to July 2020; 249 images were from patients with carpal tunnel syndrome, and 153 were from healthy volunteers. From these, 320 cases were selected as training sets, and 82 cases were selected as test sets. The improved U2-Net model was used to segment each image. Dice coefficients (Dice), pixel accuracy (PA), mean intersection over union (MIoU) and average Hausdorff distance (AVD) were used to evaluate segmentation performance. Results revealed that the Dice, MIoU, PA and AVD values of our improved U2-Net were 72.85%, 79.66%, 95.92% and 51.37 mm, respectively, which were comparable to the actual ground truth; the ground truth came from the labeling of clinicians. However, the Dice, MIoU, PA and AVD values of U-Net were 43.19%, 65.57%, 86.22% and 74.82 mm, and those of Res-U-Net were 58.65%, 72.53%, 88.98% and 57.30 mm. Overall, our data suggest our improved U2-Net model might be used for segmentation of ultrasound median neural images.

      Key Words

      Introduction

      Carpal tunnel syndrome (CTS) is a disease in which the median nerve is compressed in the carpal tunnel, disrupting sensation and function in the innervation area (
      • Pempel D
      • Evanoff B
      • Amadio PC
      • de Krom M
      • Franklin G
      • Franzblau A
      • Gray R
      • Gerr F
      • Hagberg M
      • Hales T
      • Katz JN
      • Pransky G.
      Consensus criteria for the classification of carpal tunnel syndrome in epidemiologic studies.
      ). It is the most common nerve compression disorder, and often occurs in women between 40 and 60 y of age (
      • Alfonso C
      • Jann S
      • Massa R
      • Torreggiani A
      Diagnosis, treatment and follow-up of the carpal tunnel syndrome: A review.
      ). The typical clinical manifestations of CTS are numbness in the radial three and a half fingers and a representative history of waking up with numbness in the middle of the night (
      • de Krom MC
      • van Croonenborg JJ
      • Blaauw G
      • Scholten RJ
      • Spaans F.
      Guideline 'Diagnosis and treatment of carpal tunnel syndrome.
      ). In recent years, ultrasound has become the main diagnostic method for CTS because of its low cost and convenience (
      • Cartwright MS
      • Hobson-Webb LD
      • Boon AJ
      • Alter KE
      • Hunt CH
      • Flores VH
      • Werner RA
      • Shook SJ
      • Thomas TD
      • Primack SJ
      • Walker FO.
      Evidence-based guideline: Neuromuscular ultrasound for the diagnosis of carpal tunnel syndrome.
      ). However, ultrasonic examination also has its disadvantages; sometimes local compression is difficult to detect, and the examination results can be subjective.
      At present, artificial intelligence has been widely used in various imaging diagnoses and treatments, and the analysis of images by computer-aided diagnosis (CAD) was first reported in the early 1960s (
      • Mendelsohn ML
      • Kolman WA
      • Perry B
      • Prewitt JM.
      Morphological analysis of cells and chromosomes by digital computer.
      ). For the past few years, deep learning has become an important research area of machine learning and has been involved in remarkable achievements in the field of computer vision. Deep learning has thus been applied to medical image segmentation, positioning, detection and image fusion, and has facilitated the rapid diagnosis of lesions (
      • Shen YT
      • Chen L
      • Yue WW
      • Xu HX.
      Artificial intelligence in ultrasound.
      ). As a new biomedical image processing technology, medical image segmentation has made major contributions to sustainable medicine and has become an important research direction in the field of computer vision (
      • Kaluarachchi T
      • Reis A
      • Nanayakkara S.
      A review of recent deep learning approaches in human-centered machine learning.
      ). Currently, image segmentation is being used to assess the breast and thyroid system, the abdomen and pelvis, obstetric heart and blood vessels, the musculoskeletal system and other organs (
      • Bargsten L
      • Raschka S
      • Schlaefer A.
      Capsule networks for segmentation of small intravascular ultrasound image datasets.
      ;
      • Lang S
      • Xu Y
      • Li L
      • Wang B
      • Yang Y
      • Xue Y
      • Shi K.
      Joint detection of Tap and CEA based on deep learning medical image segmentation: Risk prediction of thyroid cancer.
      ;
      • Lee K
      • Kim JY
      • Lee MH
      • Choi CH
      • Hwang JY.
      Imbalanced loss-integrated deep-learning-based ultrasound image analysis for diagnosis of rotator-cuff tear.
      ;
      • Lian J
      • Zhang M
      • Jiang N
      • Bi W
      • Dong X.
      Feature extraction of kidney tissue image based on ultrasound image segmentation.
      ;
      • Wang K
      • Liang S
      • Zhong S
      • Feng Q
      • Ning Z
      • Zhang Y.
      Breast ultrasound image segmentation: A coarse-to-fine fusion convolutional neural network.
      ,
      • Wang Z
      • Zou Y
      • Liu PX.
      Hybrid dilation and attention residual U-Net for medical image segmentation.
      ;
      • Zeng Y
      • Tsui PH
      • Wu W
      • Zhou Z
      • Wu S.
      Fetal ultrasound image segmentation for automatic head circumference biometry using deeply supervised attention-gated V-Net.
      ;
      • Zhuang S
      • Li F
      • Raj ANJ
      • Ding W
      • Zhou W
      • Zhuang Z.
      Automatic segmentation for ultrasound image of carotid intimal–media based on improved superpixel generation algorithm and fractal theory.
      ). With improvements in resolution, an increase in the utilization of ultrasound in musculoskeletal system imaging has been observed; however, the correct segmentation of neural ultrasound images has remained a challenge (
      • Shin Y
      • Yang J
      • Lee YH
      • Kim S.
      Artificial intelligence in musculoskeletal ultrasound imaging.
      ). Automatic detection of median nerve structure from medical images is a key step in early diagnosis of carpal tunnel syndrome. Proper segmentation of median nerve provides clinicians with information on the mechanism, diagnosis and treatment of carpal tunnel syndrome. Correct identification of the median nerve is the first step in the diagnosis of CTS. For beginners, it is very time consuming to identify the correct median nerve image, because it is difficult to distinguish the median nerve from the flexor tendon of the finger. For experienced operators, it is time consuming to measure the median nerve. An artificial intelligence-based segmentation algorithm allows accurate localization of the median nerve and faster and more standardized measurements of the median nerve, eliminating operator dependence and inter-observer variability in standardized measurements.
      Algorithm updates, the strengthening of computing power and the availability of large-scale data are three factors that have facilitated the continuous improvement of image segmentation. At present, U-Net networks are used mainly for image segmentation, and various other networks based on U-Net networks have been proposed (
      • Shelhamer E
      • Long J
      • Darrell T.
      Fully Convolutional networks for semantic segmentation.
      ).
      • Fang L
      • Zhang L
      • Yao Y.
      Integrating a learned probabilistic model with energy functional for ultrasound image segmentation.
      proposed a new learning conceptual model, the generalized linear model (GLM), which effectively overcomes the influence of poor quality and further improves the accuracy of segmentation.
      • Su R
      • Zhang D
      • Liu J
      • MSU-Net Cheng C.
      Multi-Scale U-Net for 2D medical image segmentation.
      proposed a multiscale U-Net (Msu-Net) for medical image segmentation, which has been found to have the best performance using different data sets (
      • Su R
      • Zhang D
      • Liu J
      • MSU-Net Cheng C.
      Multi-Scale U-Net for 2D medical image segmentation.
      ). Traditional image segmentation methods (U-Net) have been used for median nerve segmentation (
      • Festen RT
      • Schrier VJMM
      • Amadio PC.
      Automated segmentation of the median nerve in the carpal tunnel using U-Net.
      ) and can be reliably used to assess median nerve size obtained by ultrasound, thus greatly reducing labor. U-Net is a convolutional neural network that was created by O. Ronneberger, P. Fischer and T. Brox. In this model, the feature maps are extracted by four subsampling and restored to original size by four upsampling, which eventually yield segmentation results. When
      • Festen RT
      • Schrier VJMM
      • Amadio PC.
      Automated segmentation of the median nerve in the carpal tunnel using U-Net.
      used the U-Net network to segment the median nerve, the Dice coefficient reached 0.88. When
      • Huang C
      • Zhou Y
      • Tan W.
      Applying deep learning in recognizing the femoral nerve block region on ultrasound images.
      used the U-Net network, its accuracy was 88.4%.
      • Horng MH
      • Yang CW
      • Sun YN
      • Yang TH.
      DeepNerve: A new convolutional neural network for the localization and segmentation of the median nerve in ultrasound image sequences.
      proposed a new coiled neural network framework based on the U-Net model to segment median nerve, and the Dice coefficient reached 0.8912, which improved segmentation performance and generated satisfactory results. Although U-Net (
      • Ibtehaz N
      • Rahman MS.
      MultiRes U-Net: Rethinking the U-Net architecture for multimodal biomedical image segmentation.
      ) networks have been used for median nerve image segmentation, the inherent noise of ultrasound images (
      • Gerritsen HJ
      • Hannan WJ
      • Ramberg EG.
      Elimination of speckle noise in holograms with redundancy.
      ;
      • George N
      • Christensen CR
      • Bennerr JS.
      Speckle noise in displays.
      ;
      • Yan JY
      • Zhuang T.
      Applying improved fast marching method to endocardial boundary detection in echocardiographic images.
      ;
      • Sites BD
      • Brull R
      • Chan VW.
      Artifacts and pitfall errors associated with ultrasound-guided regional anesthesia: Part I. Understanding the basic principles of ultrasound physics and machine operations.
      ;
      • Fu J
      • Liu J
      • Tian H.
      Dual attention network for scene segmentation.
      ;
      • Pissas T
      • Bloch E
      • Cardoso M.
      Deep iterative vessel segmentation in OCT angiography.
      ) greatly impedes the clinician's ability to distinguish carpal tunnel syndrome and segment and measure the median nerve. In addition to median neural segmentation based on a neural network, there are also many non-neural-network algorithms, such as threshold-based methods (
      • Rodrigues PS
      • Giraldi GA.
      Improving the non-extensive medical image segmentation based on Tsallis entropy.
      ), which consider only gray statistics without considering spatial location information. The graph-based method (
      • Huang QH
      • Lee SY
      • Liu LZ
      • Lu MH
      • Jin LW
      • Li AH.
      A robust graph-based segmentation method for breast tumors in ultrasound images.
      ) requires the operator to have rich inspection experience; the active contour model (
      • Huang YL
      • Jiang YR
      • Chen DR
      • Moon WK.
      Level set contouring for breast tumor in sonography.
      ) easily reduces segmentation accuracy because of the false edge and image noise of the ultrasound image. Because of the specific characteristics of ultrasound such as attenuation, penetration, uniformity, shadow, real time and operator dependence, as well as different image characteristics on different devices, it is necessary to develop a rapid, accurate and automated screening tool to identify the median nerve. The purpose of this study was to retrospectively study the role of the U2-Net model in median neural image segmentation.

      Methods

      Data set

      From October 2018 to July 2020, a total of 249 patients with carpal tunnel syndrome diagnosed by electromyography and operated on by the Department of Hand Surgery, Huashan Hospital, Fudan University, were selected, including 51 men and 198 women ranging in age from 38 to 79 y (average: 55.39 ± 7.98 y). Another 153 healthy volunteers who visited during that same period also participated, including 45 men and 108 women, ranging in age from 54 to 66 y (average: 56.20 ± 6.18 y) (patient demographics are summarized in Table 1). A total of 402 cases were included in our study. In addition, we cropped all images to 400 × 300 to eliminate unnecessary information such as device ID and acquisition time. Approximately 320 and 82 cases were randomly used as the training and test sets.
      Table 1Patient demographics
      GroupNumberAge (y)Sex (M/F)Education (y)
      Case24955.39 ± 7.9851/19815.00 ± 3.00
      Volunteer15356.20 ± 6.1845/10815.60 ± 3.00
      An experienced ultrasound practitioner used an acoustic (EPIQ5 diagnostic ultrasound scanner, Philips, Bothell, WA, USA) S15-4 linear array probe to record a transverse 2-D ultrasound image of the median nerve at the level of the peas bone and delineate the nerve boundary in the 2-D ultrasound image. This retrospective study was approved by the medical ethics committee of Huashan Hospital, Fudan University, and informed consent was obtained from each patient.

      Pre-processing algorithms

      Histogram equalization

      In this study, we applied contrast-limited adaptive histogram equalization (CLAHE) (
      • Pizer SM
      • Amburn EP
      • Austin JD.
      Adaptive histogram equalization and its variations.
      ) and a 2-D filter to enhance the image contrast. The CLAHE algorithm comprises the following steps:
      Step 1: Transform the original ultrasonic images into gray-scale images B and divide B into several image blocks β. The size of each β is k×k. The following equations were used:
      scale=255k×k


      thresholdη=max(1,limit×scale256)


      Step 2: Compute histograms for each image block.
      Step 3: Count the number of pixels θ that exceed the threshold η for each gray level of every sub-block histogram.
      Step 4: Define the cumulative histogram of each subblock as φ;φ=θ×scale.
      Step 5: Iterate over each point of the original image and calculate the φ values of the four vertices of the sub-block where the point is located; then perform the bilinear interpolation to achieve the gray value of the transformed point. The medical image before and after applying CLAHE are provided in Figure 1.
      Fig 1
      Fig. 1(a) Original image. (b) Contrast limited adaptive histogram equalization (CLAHE) image. (c) Image after 2-D filter. The arrows indicate the carpal tunnel syndrome lesions.

      2-D filter

      The 2-D filter uses a 3×3 convolution to check the original image for convolution operation, and the value of center pixel R5 = R1G1 + R2G2 + R3G3 + R4G4 + R5G5 + R6G6 + R7G7 + R8G8 + R9G9, where R denotes the 3×3 image region, and G is the convolution kernel. The convolution operation can eliminate the noise caused by the CLAHE algorithm and smooth the image.

      U2-Net

      In this study, we modified U2-Net (
      • Qin X
      • Zhang Z
      • Huang C.
      U2-Net: Going deeper with nested U-structure for salient object detection.
      ), which was originally used for target detection, and combined this with the Adam optimizer to form a modified U2-Net. As illustrated in Figure 2, our model is a two-layer nested U-shaped structure, and the outer layer is a large U-shaped structure consisting of 11 layers. Each layer is a small U-shaped structure L-RSU (L is the number of layers in the encoder, Cin and Cout denote input and output channels, respectively, and M denotes the number of channels in the internal layers of L-RSU). It includes six layers of encoding, five layers of decoding and a feature fusion module. RSU can obtain local and global contextual information, which is important for segmentation tasks.
      We applied the L-RSU module in stages 1–4 and replaced the pooling operations in the RSU module with dilated convolutions in stages 5 and 6, in which the feature maps of the RSU module can be of the same size to prevent the loss of important information. Each decoding layer has a structure similar to that of the corresponding encoding layer and receives the feature map of the coding layer as input. The feature fusion module was used to generate segmentation results. The fusion module resamples the six masks generated from the right side of the model into a uniform input size through the 1×1 convolution and generates the final probability mapRfuse.
      In addition, it is essential to select an optimizer to update the parameters in our networks. Stochastic gradient descent (SGD) is the most popular optimizer for its generalization. In this study, we used the Adam optimizer to accelerate the convergence.
      The loss function of U2-Net is based on the Dice similarity coefficient (DSC), which is defined as
      L=m=1Mwpart(m)part(m)+wfusefuse
      (1)


      where lpart(m)(M = 5) is the loss of the side output mask map Rpart(m) and AptCommandmathcallfuse is the loss of the fusion results Rfuse. wpart(m) and wfuse are the weights of each loss l. We used the DSC to calculate loss
      =12i=1Npiqi+εi=1Npi2+i=1Nqi2+ε
      (2)


      where N is the number of total pixels on the segmentation output; pi and qi are the pixels of predicted segmentation results and the ground truth label, respectively; and ε is a small value to prevent the denominator from being zero. Figure 3 illustrates the loss during training through 250 training epochs.
      Fig 3
      Fig. 3Train loss through 250 training epochs.

      Implementation details

      The iteration of the segmentation network is 250, and the batch size is 8. We used the Adam optimizer to update the gradient with an initial learning rate of 0.0001 and a momentum β1 = 0.9. The experiment was deployed under the Pytorch framework and trained with NVIDIA Titan GPU.

      Evaluation metrics

      Four evaluation metrics were used to evaluate the performance of the segmentation network in this study: Dice, PA, MIoU and AVD (
      • Taha AA
      • Hanbury A.
      Metrics for evaluating 3D medical image segmentation: Analysis, selection, and tool.
      ). Among these, Dice is the most frequently used metric in validating medical pixel segmentations, and measures the similarity between the ground truth (the ground truth boundary comes from the labeling of clinicians, and we selected three clinicians for labeling, including two for labeling and one for auditing) and segmentation result and is defined as
      Dice=2i=1Npiqii=1Npi2+i=1Nqi2
      (3)


      where N is the number of total pixels on the prediction masks; and pi and qi are the pixels of the predicted segmentation result and the ground truth label, respectively. PA is given by the expression
      PA=TP+TNTP+TN+FP+FN
      (4)


      where TP, TN, FP and FN represent the amounts of true-positive, true-negative, false-positive, and false-negative results, respectively.
      MIoU is the ratio of the intersection and union of predicted results and the ground truth, and can be expressed as
      MIoU=TPFN+FP+TP
      (5)


      AVD describes the degree of similarity between two sets of points and is defined as
      {H(A,B)=max(h(A,B),h(B,A))h(A,B)=max(aA)min(bεB)abh(B,A)=max(bB)min(aεA)ba
      (6)


      where A and B are the sets of predicted results and the ground truth, a and b are members of sets A and B and · is the Euclidean distance. The lower the AVD value, the closer is the segmentation result to ground truth.

      Statistical analysis

      We used analysis of variance (ANOVA) tests and t-tests to verify whether the segmentation performance of our proposed method is significantly different from that of the other networks. These tests were statistically analyzed using Python 3.8 and GraphPad Prism 8.

      Results

      Because previous studies did not perform CTS segmentation, to verify the superiority of our model we compared it with classical segmentation networks such as U-Net (
      • Ronneberger O
      • Fischer P
      • Brox T
      U-Net: Convolutional networks for biomedical image segmentation.
      ) and Res-U-Net (
      • Xiao X
      • Lian S
      • Luo Z.
      Weighted Res-UNet for high-quality retina vessel segmentation.
      ). Of these, U-Net is a semantic segmentation algorithm using fully convolutional networks. Res-U-Net is based on the original U-Net model and adds a weighted attention mechanism.
      As outlined in Table 2, when compared with the other networks, our proposed modified U2-Net achieved the best segmentation performance values, including Dice (72.85%), the best MIoU (74.36%), the best PA (87.92%) and the best AVD (113.65 mm), which are closest to the ground truth. In addition, the metrics of Res-U-Net were significantly better than those of U-Net. These findings indicate that local and global contextual information is more important for segmentation tasks.
      Table 2Evaluation metrics for different segmentation networks
      Segmentation networkMetric
      Dice (%)MIoU (%)PA (%)AVD (mm)
      U-Net43.19±0.3165.57± 0.1586.22±0.1974.82 ± 66.49
      Res-U-Net58.65±0.2772.53±0.1388.98±0.1257.30±55.54
      U2-Net72.85±0.2579.66±0.1395.92±0.0151.37±49.54
      AVD = average Hausdorff distance; Dice = Dice coefficient; MIoU = mean intersection over union; PA = pixel accuracy.
      In Figure 4 are the boxplots of the different networks’ evaluation metrics. The boxplot is a statistical chart that depicts the distribution of data, including the maximum, minimum, median and upper and lower quartiles of a set of data. The results of U-Net and Res-U-Net reveal a wide range of distribution, which indicates their test results are largely influenced by the test samples. In contrast, the results of the modified U2-Net are more concentrated and have the best data distribution results, which indicates that our model is more robust. As outlined in Table 3, the p values of ANOVAs and t-tests are all <0.05, which indicates that our model has a significantly better performance.
      Fig 4
      Fig. 4Boxplots of different networks’ evaluation metrics. AVD = average Hausdorff distance; Dice = Dice coefficient; MIoU = mean intersection over union; PA = pixel accuracy.
      Table 3ANOVA test and t-test on metrics for different segmentation networks
      U2-Net vs. U-netU2-Net vs. Res-U-Net
      p ValueLevene's testANOVAt-TestLevene's testANOVAt-Test
      Dice0.2350.0270.0380.0790.0160.005
      MIoU0.0580.0030.0010.3080.0050.007
      AVD0.7650.0140.0310.4730.0310.026
      PA0.1260.0390.0020.0750.0240.039
      ANOVA = analysis of variance; AVD = average Hausdorff distance; Dice = Dice coefficient; MIoU = mean intersection over union; PA = pixel accuracy.
      The segmentation results of the different networks are illustrated in Figure 5. The (a) U-Net and (b) Res-U-Net misclassified areas that were not lesions in case 1 and divided the region of interest into background sections in case 2 (c). U2-Net performs well in both cases 1 and 2, which exhibits minimal undersegmentation and is closer to the ground truth. In case 3, the three models cannot segment the region of interest because of the extreme similarity between background and foreground, but U2-Net is more stable in various ultrasound cases compared with the other models.
      Fig 5
      Fig. 5Columns from left to right are the segmentation results of carpal tunnel syndrome lesions by different networks and the ground truth. (a) Original image. (b) U-Net. (c) Res-U-Net. (d) U2-Net. Rows from up to down are the samples (cases 1–3).

      Discussion

      Our study describes a new image segmentation method that first uses CLAHE and a 2-D filter to enhance image contrast, then uses coil weaving to eliminate noise generated by the CLAHE algorithm to generate smoother images and combines this with the Adam optimizer to optimize the algorithm to form an improved U2-Net model. As outlined in Table 1, the performance of U2-Net is significantly better than that of the U-Net and Res-U-Net models. As illustrated in Figure 4, the U2-Net model mitigates the disadvantage of U-Net's and Res-U-Net's ability to be influenced by test samples, and its results are denser and generate the best data distribution. These results indicate that the improved U2-Net model used in this study enhances the accuracy of median nerve segmentation.
      The U-Net model is one of the most popular whole rolls of woven network models and has been widely used in medical image segmentation, including both the encoder and the decoder, between which layers are skipped to connect the two (
      • Ronneberger O
      • Fischer P
      • Brox T
      U-Net: Convolutional networks for biomedical image segmentation.
      ). It can extract image features with limited samples and achieve high segmentation accuracy, exceeding the accuracy of traditional methods, and it illustrates the potential of efficient automatic segmentation. It can reduce delineation time and eliminate differences between and within observers (
      • Young AV
      • Wortham A
      • Wernick I
      • Ennis RD.
      Atlas-based segmentation improves consistency and decreases time required for contouring postoperative endometrial cancer nodal volumes.
      ;
      • Daisne JF
      • Blumhofer A.
      Atlas-based automatic segmentation of head and neck organs at risk and nodal target volumes: A clinical validation.
      ). At present, there are many variant networks of U-Net such as Res-U-Net, HDA-Res-U-Net (
      • Wang K
      • Liang S
      • Zhong S
      • Feng Q
      • Ning Z
      • Zhang Y.
      Breast ultrasound image segmentation: A coarse-to-fine fusion convolutional neural network.
      ,
      • Wang Z
      • Zou Y
      • Liu PX.
      Hybrid dilation and attention residual U-Net for medical image segmentation.
      ), CS2-Net (
      • Mou L
      • Zhao Y
      • Fu H
      • Liu Y
      • Cheng J
      • Zheng Y
      • Su P
      • Yang J
      • Chen L
      • Frangi AF
      • Akiba M
      • Liu J.
      CS2-Net: Deep learning segmentation of curvilinear structures in medical imaging.
      ) and UV-Net (
      • Zhang C
      • Hua Q
      • Chu Y
      • Wang P.
      Liver tumor segmentation using 2.5D UV-Net with multi-scale convolution.
      ), all of which have achieved good segmentation results. U-Net can be realized in 2-D or 3-D format, both of which have their advantages and disadvantages. With 2-D U-Net, numerous samples can be learned, but because each image is independently processed, the 3-D direction of information is reduced. However, while using 3-D U-Net, the number of samples is decreased, but the amount of information of each sample is increased, and so the 3-D information direction is enriched. In addition, compared with 2-D U-Net, 3-D U-Net requires a larger amount of computing resources and longer computing time (
      • Nemoto T
      • Futakami N
      • Yagi M
      • Kumabe A
      • Takeda A
      • Kunieda E
      • Shigematsu N.
      Efficacy evaluation of 2D, 3D U-Net semantic segmentation and atlas-based segmentation of normal lungs excluding the trachea and main bronchi.
      ). In this study, 2-D U-Net was adopted to train through slices and make full use of 2-D plane spatial information in each slice. In terms of the algorithm, parallel coiling cores of different sizes are used to view the image region of interest of different scales in the initial block. Currently, 1 × 1, 3 × 3, 5 × 5 and 7 × 7 coiling operations are commonly used. The 1 × 1 coiling operation can reduce the number of parameters and broaden the network channel with the least parameters. In this study, we used 3 × 3 volume weaving to reduce computational overhead and network parameters. We obtained output information from three volume-weaving blocks, connected them together to extract spatial features from different scales and generated good results.
      The model we proposed not only retains the characteristics of traditional U-Net, but also adds the L-RSU module to learn low-dimensional and high-dimensional features. The convergence of the loss function can be accelerated by calculating the difference between the segmentation results and the ground truth of each layer. The models proposed by
      • Festen RT
      • Schrier VJMM
      • Amadio PC.
      Automated segmentation of the median nerve in the carpal tunnel using U-Net.
      and
      • Horng MH
      • Yang CW
      • Sun YN
      • Yang TH.
      DeepNerve: A new convolutional neural network for the localization and segmentation of the median nerve in ultrasound image sequences.
      were simply applications for U-Net. In addition, our data set contains only 402 images, whereas those of Festen et al. and Horng et al. contained 5560 and 1680 images, respectively. But our Dice coefficient is only about 0.1 lower than that for their model. Therefore, our proposed model will definitely perform better than their algorithms if we have the same amount of data.
      Compared with manual and other automatic segmentation methods, the U2-Net model does have a performance advantage. However, there is room for improvement in terms of accuracy based on some limitations: (i) Our accuracy is based on the manual segmentation, which is influenced by the user's variability and change. (ii) The 3 × 3 volume weaving operation leads to relatively few-feature scales and a lack of multiscale features. (iii) Because U2-Net uses many slices but ignores the information between slices, the segmentation accuracy is reduced. (iv) The design of the image segmentation network is usually severely limited by central processing unit and computer memory. (v) The sample size of the test is limited, and the segmentation of medical images requires more annotated high-resolution images, which requires individuals with medical backgrounds to annotate a large number of training image data sets. (vi) This study only examined only the median nerve, and thus excluded other peripheral and small nerves. (vii) This study could only identify the median nerve, not diagnose disease, and was unable to determine disease severity.
      Ultrasound has been extensively used in routine peripheral nerve examination. Electrophysiological examination and magnetic resonance examination are also two important examination methods. Electrophysiological examination is considered the gold standard for the diagnosis of peripheral neuropathy, but it can cause trauma to the patient and may generate false-negative results. Although magnetic resonance imaging is currently the best imaging test for diagnosing peripheral neuropathy, it is expensive and unsuitable for patients with metal implants and claustrophobia. Manual segmentation is generally time consuming and laborious, and thus the demand for automatic neural segmentation has increased. Automatic or semi-automatic artificial intelligence–based segmentation exhibits potential advantages with respect to peripheral nerve segmentation because of its nearly instant evaluation, cost-effectiveness and high reproducibility. Many image segmentation models have been used to segment peripheral nerves with good results.

      Conclusions

      Image segmentation has been widely used in ultrasound. The main contribution of this study was development of an improved model called U2-Net, which was compared with U-Net and Res-U-Net. The results indicate that the segmentation accuracy of U2-Net is superior, thus providing evidence that our model may be interchanged with manual segmentation. In view of the good performance of U2-Net in median nerve segmentation, it would be productive to further study these methods and improve them for potential use in the image segmentation of other neural entities.

      Conflict of interest disclosure

      The authors declare that they have no conflicts of interest.

      References

        • Alfonso C
        • Jann S
        • Massa R
        • Torreggiani A
        Diagnosis, treatment and follow-up of the carpal tunnel syndrome: A review.
        Neurol Sci. 2010; 31: 243-252
        • Bargsten L
        • Raschka S
        • Schlaefer A.
        Capsule networks for segmentation of small intravascular ultrasound image datasets.
        Int J Comput Assist Radiol Surg. 2021; 7: 1861-1872
        • Cartwright MS
        • Hobson-Webb LD
        • Boon AJ
        • Alter KE
        • Hunt CH
        • Flores VH
        • Werner RA
        • Shook SJ
        • Thomas TD
        • Primack SJ
        • Walker FO.
        Evidence-based guideline: Neuromuscular ultrasound for the diagnosis of carpal tunnel syndrome.
        Muscle Nerve. 2012; 46: 287-293
        • Daisne JF
        • Blumhofer A.
        Atlas-based automatic segmentation of head and neck organs at risk and nodal target volumes: A clinical validation.
        Radiat Oncol. 2013; 154 (1748–1717)
        • de Krom MC
        • van Croonenborg JJ
        • Blaauw G
        • Scholten RJ
        • Spaans F.
        Guideline 'Diagnosis and treatment of carpal tunnel syndrome.
        Ned Tijdschr Geneeskd. 2008; 152: 76-81
        • Fang L
        • Zhang L
        • Yao Y.
        Integrating a learned probabilistic model with energy functional for ultrasound image segmentation.
        Med Biol Eng Comput. 2021; 59: 1917-1931
        • Festen RT
        • Schrier VJMM
        • Amadio PC.
        Automated segmentation of the median nerve in the carpal tunnel using U-Net.
        Ultrasound Med Biol. 2021; 47: 1964-1969
        • Fu J
        • Liu J
        • Tian H.
        Dual attention network for scene segmentation.
        in: Proceedings, 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 32. IEEE, New York2019: 3146-3154 (Long Beach, CA, June 15–20)
        • George N
        • Christensen CR
        • Bennerr JS.
        Speckle noise in displays.
        J Opt Soc Am. 1976; 66: 1282-1290
        • Gerritsen HJ
        • Hannan WJ
        • Ramberg EG.
        Elimination of speckle noise in holograms with redundancy.
        Appl Opt. 1968; 7: 2301-2311
        • Horng MH
        • Yang CW
        • Sun YN
        • Yang TH.
        DeepNerve: A new convolutional neural network for the localization and segmentation of the median nerve in ultrasound image sequences.
        Ultrasound Med Biol. 2020; 46: 2439-2452
        • Huang YL
        • Jiang YR
        • Chen DR
        • Moon WK.
        Level set contouring for breast tumor in sonography.
        J Digit Imaging. 2007; 20: 238-247
        • Huang QH
        • Lee SY
        • Liu LZ
        • Lu MH
        • Jin LW
        • Li AH.
        A robust graph-based segmentation method for breast tumors in ultrasound images.
        Ultrasonics. 2012; 52: 266-275
        • Huang C
        • Zhou Y
        • Tan W.
        Applying deep learning in recognizing the femoral nerve block region on ultrasound images.
        Ann Transl Med. 2019; 7: 453-460
        • Ibtehaz N
        • Rahman MS.
        MultiRes U-Net: Rethinking the U-Net architecture for multimodal biomedical image segmentation.
        Neural Netw. 2020; 121: 74-87
        • Kaluarachchi T
        • Reis A
        • Nanayakkara S.
        A review of recent deep learning approaches in human-centered machine learning.
        Sensors (Basel). 2021; 21: 2514-2543
        • Lang S
        • Xu Y
        • Li L
        • Wang B
        • Yang Y
        • Xue Y
        • Shi K.
        Joint detection of Tap and CEA based on deep learning medical image segmentation: Risk prediction of thyroid cancer.
        J Healthc Eng. 2021; 6: 1-9
        • Lee K
        • Kim JY
        • Lee MH
        • Choi CH
        • Hwang JY.
        Imbalanced loss-integrated deep-learning-based ultrasound image analysis for diagnosis of rotator-cuff tear.
        Sensors (Basel). 2021; 21: 2214-2228
        • Lian J
        • Zhang M
        • Jiang N
        • Bi W
        • Dong X.
        Feature extraction of kidney tissue image based on ultrasound image segmentation.
        J Healthc Eng. 2021; 4: 1155-1171
        • Mendelsohn ML
        • Kolman WA
        • Perry B
        • Prewitt JM.
        Morphological analysis of cells and chromosomes by digital computer.
        Methods Inf Med. 1965; 4: 163-167
        • Mou L
        • Zhao Y
        • Fu H
        • Liu Y
        • Cheng J
        • Zheng Y
        • Su P
        • Yang J
        • Chen L
        • Frangi AF
        • Akiba M
        • Liu J.
        CS2-Net: Deep learning segmentation of curvilinear structures in medical imaging.
        Med Image Anal. 2021; 67101874
        • Nemoto T
        • Futakami N
        • Yagi M
        • Kumabe A
        • Takeda A
        • Kunieda E
        • Shigematsu N.
        Efficacy evaluation of 2D, 3D U-Net semantic segmentation and atlas-based segmentation of normal lungs excluding the trachea and main bronchi.
        J Radiat Res. 2020; 61: 257-264
        • Pempel D
        • Evanoff B
        • Amadio PC
        • de Krom M
        • Franklin G
        • Franzblau A
        • Gray R
        • Gerr F
        • Hagberg M
        • Hales T
        • Katz JN
        • Pransky G.
        Consensus criteria for the classification of carpal tunnel syndrome in epidemiologic studies.
        Am J Public Health. 1998; 88: 1447-1451
        • Pissas T
        • Bloch E
        • Cardoso M.
        Deep iterative vessel segmentation in OCT angiography.
        Biomed Opt Express. 2020; 11: 2490-2510
        • Pizer SM
        • Amburn EP
        • Austin JD.
        Adaptive histogram equalization and its variations.
        Comput Vis Graph Image Process. 1987; 39: 355-368
        • Qin X
        • Zhang Z
        • Huang C.
        U2-Net: Going deeper with nested U-structure for salient object detection.
        Pattern Recognition. 2020; 106107404
        • Rodrigues PS
        • Giraldi GA.
        Improving the non-extensive medical image segmentation based on Tsallis entropy.
        Pattern Anal Appl. 2011; 14: 369-379
        • Ronneberger O
        • Fischer P
        • Brox T
        U-Net: Convolutional networks for biomedical image segmentation.
        in: Navab N Hornegger J Wells W Frangi A Medical image computing and computer-assisted intervention—MICCAI 2015. Lecture Notes in Computer Science. 9351. Springer, Cham2015: 234-241 (Vol.)
        • Shelhamer E
        • Long J
        • Darrell T.
        Fully Convolutional networks for semantic segmentation.
        Proc IEEE Comput Soc Conf Comput Vis Pattern Recognit. 2017; 39: 640-651
        • Shen YT
        • Chen L
        • Yue WW
        • Xu HX.
        Artificial intelligence in ultrasound.
        Eur J Radiol. 2021; 139109717
        • Shin Y
        • Yang J
        • Lee YH
        • Kim S.
        Artificial intelligence in musculoskeletal ultrasound imaging.
        Ultrasonography. 2021; 40: 30-44
        • Sites BD
        • Brull R
        • Chan VW.
        Artifacts and pitfall errors associated with ultrasound-guided regional anesthesia: Part I. Understanding the basic principles of ultrasound physics and machine operations.
        Reg Anesth Pain Med. 2007; 32: 412-418
        • Su R
        • Zhang D
        • Liu J
        • MSU-Net Cheng C.
        Multi-Scale U-Net for 2D medical image segmentation.
        Front Genet. 2021; 12: 63993
        • Taha AA
        • Hanbury A.
        Metrics for evaluating 3D medical image segmentation: Analysis, selection, and tool.
        BMC Med Imaging. 2015; 15: 29-35
        • Wang K
        • Liang S
        • Zhong S
        • Feng Q
        • Ning Z
        • Zhang Y.
        Breast ultrasound image segmentation: A coarse-to-fine fusion convolutional neural network.
        Med Phys. 2021; 3: 2405-2435
        • Wang Z
        • Zou Y
        • Liu PX.
        Hybrid dilation and attention residual U-Net for medical image segmentation.
        Comput Biol Med. 2021; 134104449
        • Xiao X
        • Lian S
        • Luo Z.
        Weighted Res-UNet for high-quality retina vessel segmentation.
        in: 2018 9th International Conference on Information Technology in Medicine and Education (ITME). IEEE, New York2018: 327-331
        • Yan JY
        • Zhuang T.
        Applying improved fast marching method to endocardial boundary detection in echocardiographic images.
        Pattern Recognit Lett. 2003; 24: 2777-2784
        • Young AV
        • Wortham A
        • Wernick I
        • Ennis RD.
        Atlas-based segmentation improves consistency and decreases time required for contouring postoperative endometrial cancer nodal volumes.
        Int J Radiat Oncol Biol Phys. 2011; 79: 943-947
        • Zeng Y
        • Tsui PH
        • Wu W
        • Zhou Z
        • Wu S.
        Fetal ultrasound image segmentation for automatic head circumference biometry using deeply supervised attention-gated V-Net.
        J Digit Imaging. 2021; 34: 134-148
        • Zhang C
        • Hua Q
        • Chu Y
        • Wang P.
        Liver tumor segmentation using 2.5D UV-Net with multi-scale convolution.
        Comput Biol Med. 2021; 133104424
        • Zhuang S
        • Li F
        • Raj ANJ
        • Ding W
        • Zhou W
        • Zhuang Z.
        Automatic segmentation for ultrasound image of carotid intimal–media based on improved superpixel generation algorithm and fractal theory.
        Comput Methods Programs Biomed. 2021; 205106084