. 2022 Jun 3;35(6):1433–1444. doi: 10.1007/s10278-022-00661-4

Deep CTS: a Deep Neural Network for Identification MRI of Carpal Tunnel Syndrome

Haiying Zhou ¹, Qi Bai ³, Xianliang Hu ³, Ahmad Alhaskawi ¹, Yanzhao Dong ¹, Zewei Wang ⁵, Binjie Qi ², Jianyong Fang ⁴, Vishnu Goutham Kota ⁵, Mohamed Hasan Abdulla Hasa Abdulla ⁵, Sohaib Hasan Abdullah Ezzi ⁵, Hui Lu ^1,^6,^✉

PMCID: PMC9712834 PMID: 35661280

Abstract

Carpal tunnel syndrome (CTS) is a common peripheral nerve disease in adults; it can cause pain, numbness, and even muscle atrophy and will adversely affect patients’ daily life and work. There are no standard diagnostic criteria that go against the early diagnosis and treatment of patients. MRI as a novel imaging technique can show the patient’s condition more objectively, and several characteristics of carpal tunnel syndrome have been found. However, various image sequences, heavy artifacts, small lesion characteristics, high volume of imagine reading, and high difficulty in MRI interpretation limit its application in clinical practice. With the development of automatic image segmentation technology, the algorithm has great potential in medical imaging. The challenge is that the segmentation target is too small, and there are two categories of images with the proximal border of the carpal tunnel as the boundary. To meet the challenge, we propose an end-to-end deep learning framework called Deep CTS to segment the carpal tunnel from the MR image. The Deep CTS consists of the shape classifier with a simple convolutional neural network and the carpal tunnel region segmentation with simplified U-Net. With the specialized structure for the carpal tunnel, Deep CTS can segment the carpal tunnel region efficiently and improve the intersection over union of results. The experimental results demonstrated that the performance of the proposed deep learning framework is better than other segmentation networks for small objects. We trained the model with 333 images, tested it with 82 images, and achieved 0.63 accuracy of intersection over union and 0.17 s segmentation efficiency, which indicate great promise for the clinical application of this algorithm.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10278-022-00661-4.

Keywords: Deep learning, Image segmentation, Carpal tunnel syndrome, MRI

Background

Carpal tunnel syndrome (CTS) is when the median nerve is compressed as it passes through the carpal tunnel. It is the most common entrapment neuropathy, and it is a condition that causes numbness, paresthesia, pain, and other symptoms in the hand [1]. Electromyography (EMG) is generally regarded as the gold standard for diagnosing CTS [2]. But it only measures the physiological function and cannot observe the anatomical pathological abnormalities, and it is invasive [3]. Magnetic resonance image (MRI) can clearly detect the structure of nerves and other soft tissues and intuitively display the volume of the carpal tunnel [4, 5]. But everything has the dual character. The detailed presentation of anatomical structures on MRI also results in multiple sequences and a large number of images, which not only makes it take longer to perform interpretations, but also small tissue lesions may be overlooked due to uncoordinated case information and too many interferences of imaging such as artifacts [6]. Meanwhile, the lack of imaging expertise and regional variability also limit the widespread use of MRI [7, 8].

In recent years, convolutional neural networks (CNNs) have achieved great success in computer vision, especially in image classification [9–12] and segmentation.

Semantic segmentation is the basic task in segmentation. Each pixel in image needs to be classified.

Fully convolutional networks are the origin of most semantic segmentation models today (FCN) [13]. It realized end-to-end image segmentation for the first time. The insight of this model was to use the existing CNNs as powerful feature extractor to extract visual features which called the spatial maps. Those maps are upsampled by transposed convolution, and pixel-level category prediction results are obtained. The aforementioned feature extraction and upsampling are also referred to as encoding and decoding. SegNet is a typical example of decoder variant [14]. The decoder of SegNet consists of a set of upsampling which uses pooling indices and convolution layers. In order to integrate the global information of the image, the DeepLab model uses dilated convolution to increase the receptive field of the convolution kernel and uses conditional random fields as a post-processing step to refine the segmentation result [15].

Instance segmentation is another category of segmentation tasks, but more difficult than semantic segmentation. It requires the correct detection of all objects in an image while also precisely segmenting each instance. Mask R-CNN combines object detection and semantic segmentation together to achieve good results [16]. Girshick et al. proposed region-based CNN to detect objects in an image [17]. This model uses selective search to locate the region proposals. Then, CNNs are used on each region proposal to get the features, which are classified by a linear support vector machine (SVM). Finally, regression is applied to the previous proposals to refine it. Fast-RCNN interchanges the order of selective search and feature extraction and adds region of interest (ROI) pooling after feature extraction [18]. In addition, Fast-RCNN uses softmax instead of SVM for classification. Faster R-CNN improves the model of Fast-RCNN [19]. The selective search in Fast-RCNN is replaced by region proposal network (RPN), which makes the overall performance greatly improved, especially in terms of detection speed. Mask R-CNN is an improvement based on the Faster R-CNN model. It adds a mask prediction branch to enable the network to complete object detection and semantic segmentation.

The deep convolutional neural network models mentioned above all require a large number of data for training. The acquisition of more data is important for good results; however, to the best of our knowledge, there are few publicly available MRI of CTS datasets. The dataset we collected ourselves is small, with only 415 samples. In order to solve the problem of small sample medical image segmentation, U-Net [20] is raised. Compared with natural images, MRIs of the carpal tunnel have mainly two categories of forms and have a simple background. In this sense, images we used in carpal tunnel segmentation task have relatively simple semantics and relatively fixed structure. With the use of data augmentation, U-Net can be trained from very few images.

In this research, we focus our attention on segmenting the carpal tunnel from the MRI image. Compared with FCN, in the upsampling part, U-Net has a large number of feature channels, which allow the network to propagate context information to higher resolution layers. Unlike SegNet, U-Net does not use pooling indices for upsampling which may lose important parts of carpal tunnel MRI. In addition, the dataset we use is small and cannot meet the requirements of general deep networks. Therefore, we use the U-Net model with minor changes. One of our obstacles is that the carpal tunnel is too small for the total area of the picture. In the dataset, the average proportion of the carpal tunnel in the original image is only 0.87%. Even after the cropping operation in data preprocessing, the proportion only increased to 1.85%. Another difficulty is that MRIs of CTS have two category shapes with the proximal border of the carpal tunnel as the boundary. To address these problems, we propose a method to achieve end-to-end carpal tunnel segmentation by first classifying the image and then segmenting it.

Materials

Patient Data

Patients who had symptoms of CTS were screened between August 2019 and October 2020. Written informed consent was obtained from the patients for publication of this research and any accompanying images. Ethical approval was given by the medical ethics committee of the First Affiliated Hospital, College of Medicine, Zhejiang University.

Preoperative symptoms, physical examination, electromyography, and intraoperative findings all confirmed the diagnosis. The patient underwent a magnetic resonance examination before the operation. All MRI examinations were performed on an MR imager (Philips Achieva 3.0 T, Netherlands). The patient was in a prone position, keeping the examined hand extended over the head in a neutral position, which subsequently reached into a coil designed for the wrist with the palm down and the examined wrist joint placed in the center of the main magnetic field, and the scan ranges from the distal radius to the proximal phalanges. The MRI parameters of conventional image scanning sequence were as follows: TR/TE = 550 ms / 12 ms (T1WI sequence), TR/TE = 2500 ms / 48 ms (PDWI sequence) 2.5-mm slice thickness, 12-cm field of view (FOV), 256 × 256 matrix. Exclusion criteria include previous wrist trauma, previous wrist surgery, carpal tunnel tumors, gout, synovial hyperplasia, and other space-occupying lesions, wrist skin and soft tissue infections, hypothyroidism, rheumatoid arthritis, pregnant women, and cervical radiculopathy. The patients were in the supine position, with their wrists in the neutral position. The detailed data of patients were recorded: age, sex, weight, height, occupation, medical history, dominant hand, affected hand, complaints, and duration of symptoms.

Description of the Dataset

This special dataset consists of MRI and corresponding carpal tunnel label data. The proximal and distal borders of the carpal tunnel (pisiform bone and hook of the hamate) are identified in the coronal plane, and the images of the median nerve with different cross-sections are divided into two categories using the proximal border of the carpal tunnel as the boundary, given that when compressed, the median nerve becomes flattened within the carpal tunnel and swollen before entering the carpal tunnel [21]. Then, the enlarged image and the median nerve label were marked by an experienced orthopedic surgeon and a junior radiologist to ensure accuracy, making the labeling criteria based on both clinical and radiologic perspectives. And in case of ambiguous marked images, these two experts would discuss to reach consensus, and eventually, the labels were all confirmed by intraoperative findings. The whole database of hand MRI consisted of 415 pairs of images and labels, including both left and right hands. There are some examples in the dataset of hand MRI, as shown in Fig. 1. The upper part of the figure is the MRIs of the hand, and the lower part is the corresponding marks, in which the carpal tunnel is marked in red. The two pairs of data on the left are in distal category. The two on the right are in proximal, and other information is carried around them.

Data Preparation

For deep learning image segmentation tasks, the quality of the samples determined the quality of the output. However, the similar shape, texture, and gray value of MRI images make the segmentation task difficult. To tackle these difficulties, we preprocess the MRI images by resizing, cropping, and Contrast Limited Adaptive Histogram Equalization (CLAHE).

The original MRIs we got are not in the same size. Most of the images are 512 × 512, but a few of the images are too small. We first use the nearest neighbor interpolation method to resize the image to a uniform size of 512 × 512, which can ensure that the model can be trained in batches with the same input size. Some information that is not useful for carpal tunnel segmentation appears at the edge of MRIs, such as gender and time. In order to remove the effect of information carried on the edge of MRIs and to increase the area of the carpal tunnel in the total image, we keep the center of the image unchanged, crop the edges of the image, and make the image to a size of 320 × 384, which could also properly obtain the main part of the image. The medical images captured from MRI technology suffer from noise and poor contrast issues. Therefore, contrast enhancement techniques are widely used to improving the quality of MRIs. Pizer et al. proposed a method called Contrast Limited Adaptive Histogram Equalization (CLAHE) to overcome the amplification of noise problem and to improve the contrast for medical images [22]. At the end of data preprocessing, CLAHE is used to improve MRI quality. We set contextual region size equals 40 × 48, which divides each input image into 64 nonoverlapping blocks and clip limits equals 0.03, which can effectively alter the contrast of the MRIs.

After the preprocess above, the quality of the original images has been improved, including the shape, texture of the image, and the boundary information. The processed images were shown in Fig. 2.

Methods

Samples

As mentioned earlier, the MRIs with different cross-sections are divided into two categories using the border of the carpal tunnel as the boundary, denoted as proximal and distal. The ratio of training set to test set is eight to two. The division of training set and test set is shown in Table 1.

Table 1.

Sample distribution

	Proximal	Distal	Total
Train	115	218	333
Test	28	54	82
Total	143	272	415

Open in a new tab

The images in distal are similar to Fig. 2a, while the images in proximal are similar to Fig. 2c.

Framework of the Data Flow

When only using U-Net for carpal tunnel segmentation, the prediction results of some samples are always unsatisfactory. Those samples prompted us to explore the deeper structure or features of the MRI images. Considering that there are two categories of image form of carpal tunnel MRI, and each category of images is very similar, we may need to train a segmentation model for each category of images separately. Furthermore, in order to achieve end-to-end carpal tunnel segmentation, we add a classification model to determine which category the image belongs to before the segmentation model. The flow chart of the proposed method was shown in Fig. 3. We used the datasets divided in Table 1 to train three models separately: the classification model and the two segmentation models corresponding to the two categories. When training the classification model, we use all images, and when training the two segmentation models, we use the proximal and distal images, respectively. During the test, for each input MRI, the classification network is used to determine which category the image belongs to, and then, the corresponding segmentation network is used to obtain the segmentation result.

Conv Block

To solve the carpal tunnel segmentation problem faster and more accurately, we built a tiny-scale model for classification and simplified the architecture of U-Net for segmentation. Firstly, we introduced the basic unit we used in all models, which was the so-called convolution block.

The purpose of convolution block is to extract deeper features without changing the size of the picture, that is, to increase or decrease the number of feature maps. The architecture is illustrated in Fig. 4. In this architecture, we just use two-dimensional (2D) convolution, batch normalization (BN), and rectified linear unit (ReLU) activation function. It takes an output channels parameter to determine the action of this block. In the encoding process, before the pooling layer, we usually set the output channels to be twice the input channels, which is determined by the input data x. On the contrary, when decoding, the output channels are set to half the input channels. The convolution block consists of the repeated application of two 3 × 3 2D convolution with padding equals 1 to make sure the shape of feature maps doesn’t change, each followed by BN and ReLU. The first 2D convolution changes output channels to what we desired, and the second one keeps channels unchanged.

Classification Model

In order to obtain more accurate carpal tunnel segmentation results, we separately trained segmentation models on MRIs of the distal and proximal. The classification model is used to determine which category an MRI falls into. The architecture of the classification model is shown in Fig. 5. Only convolution block, 8 × 8 max pooling, and fully connected layer are used. It consists of a feature extraction part and a classification part. Two Conv Blocks mentioned above make up the feature extraction part, each followed by an 8 × 8 max pooling operation with stride 2, and their output channels are 16 and 32 in order. The classification part follows the typical architecture of classification task. First, the features we obtained with feature extraction part are flattened into one dimension. After that, two fully connection layers act on the features in turn, reducing the number of features to 64 and then to 2, corresponding to the two categories, respectively. The size of the model parameters is only 311 kb, so we called it a small-scale model. In the training process, the two-class cross-entropy loss function is used. Its expression is as follows:

L (y, \hat{y}) = - (y log \hat{y} + (1 - y) log (1 - \hat{y}))

where y is the sample label and $\hat{y}$ is the predicted label.

U-Net

Like other medical image segmentation tasks, we used U-Net to segment the carpal tunnel area. The simplified U-Net architecture was illustrated in Fig. 6. We only reduced the number of feature maps to half of the original U-Net.

U-Net consisted of two parts, down-sampling (left) and upsampling (right). In Fig. 6, the red downward arrows represented the result of 2 × 2 max pooling with stride 2, the yellow upward arrows represented the 2 × 2 transposed convolution which doubled the image size and made the number of feature maps half of the input, while the blue right arrows represented the result of the Conv Block. The light blue arrows represented skip connections, stitching together feature maps of the same scale.

We regarded the carpal tunnel segmentation task as a binary classification task of pixels, so we used dice loss [23] as loss function. Its expression is as follows:

L_{dice} (X, Y) = 1 - \frac{2 |X \cap Y| + ε}{|X| + |+ ε|}

where X is label, Y is prediction, and $|X|$ and $|Y|$ represent the number of pixels in X and Y, respectively. $|X \cap Y|$ represents the number of pixels in the intersection of X and Y. $ε$ is a sufficiently small positive number to avoid zero denominator.

Training and Evaluation Criteria

In this work, we implemented our model using PyTorch on four NVIDIA Tesla T4 GPU. For all models, we used Adam optimizer to optimize the loss of function and set the initial learning rate to 0.001 and batch size to 16. The classification model was trained for 40 epochs, and the learning rate decayed by 0.1 after 20 epochs. On the contrary, the two segmentation models were trained for 240 epochs, and the learning rate decayed by 0.1 after 180 epochs.

During the model training, we use fivefold cross-validation to evaluate the model. For binary classification, classes 1 and 2 are called positive and negative sample, respectively. True positive (TP) represents the number of positive samples that are correctly classified; true negative (TN) represents the number of negative samples that are correctly classified; false positive (FP) represents the number of samples originally belonging to positive but classified as negative; and false negative (FN) represents the number of samples originally belonging to negative but classified as positive. We use accuracy to evaluate the classification model. The accuracy is defined as the percentage of correct classification, which is:

Accuracy = \frac{T P + T N}{T P + F P + T N + F N}

When evaluating the segmentation models, pixel accuracy (PA) and intersection over union (IoU) of target are used. PA is defined as the percentage of correctly predicted pixels in the segmentation target. We set segmentation target as positive samples; then calculation formula of PA is:

P A = \frac{TP}{T P + F P}

IoU is defined as the ratio of the number of intersection and union pixels between predicted pixels and ground truth pixels. The IoU calculation formula is as follows:

I o U = \frac{TP}{T P + F P + F N}

The larger the accuracy, PA, and IoU, the better the performance of the corresponding model.

Experimental Results

In this article, we implement U-Net and Deep CTS. Visualization and quantitative results are used to analyze and compare the models’ performances. The segmentation quality is measured by PA and IoU, and the classification model is evaluated by accuracy. Table 2 shows the numerical indicators of models.

Table 2.

Numerical indicators

	Criteria	Fold 1	Fold 2	Fold 3	Fold 4	Fold 5	Average
U-Net	PA	0.4666	0.6238	0.7093	0.6494	0.7105	0.6319
U-Net	IoU	0.4113	0.5324	0.5381	0.5486	0.5901	0.5241
Classification	Acc	0.9634	0.9756	0.9390	1.000	0.9756	0.9707
Deep CTS	PA	0.7445	0.7644	0.7073	0.7607	0.7524	0.7459
Deep CTS	IoU	0.6158	0.6565	0.6087	0.6253	0.6487	0.6310

Open in a new tab

The existing U-Net model is used to conduct experiment on carpal tunnel data firstly, which achieved only 63.19% PA and 52.41% IoU on average in fivefold cross-validation. This experiment did not yield satisfactory results and indirectly demonstrated that MRIs of CTS differ from natural images. Therefore, in order to better segment the carpal tunnel, we designed the Deep CTS.

With the proximal border of the carpal tunnel as the boundary, we divide MRIs into two categories, proximal and distal. The classification model used to determine the MRI class in Deep CTS is first trained. The change of classification accuracy during training is shown in Fig. 7. The model is simple, but it has achieved good results in classification tasks. The accuracy rate on the test set reached 97.07% on average as shown in Table 2, which is sufficient for the classification process of Deep CTS.

Afterwards, the segmentation models for the proximal and distal are trained separately. The loss change of the segmentation models is shown in Fig. 8. Due to the uneven number of samples in the two categories, the convergence speed of the model under the same parameters is different. During the first few epochs, the value of the loss function decreased rapidly, and in the subsequent training process, the value of the loss function decreased slowly until convergence.

During the test, according to the architecture shown in Fig. 3, the three models are formed into an end-to-end carpal tunnel segmentation model called Deep CTS. Each MRI input to Deep CTS is first classified. According to the classification result, the MRI is input into the corresponding segmentation network to obtain the segmentation result. The results of Deep CTS of fivefold cross-validation are shown in Table 2; it achieved 74.59% PA and 63.10% IoU on average. Compared with U-Net, Deep CTS improves PA by 11.40% and IoU by 10.69%. The performance of the model has been greatly improved, on the one hand, because we have trained segmentation models for the proximal and distal categories respectively, so that the models can better capture the unique features of the two categories. On the other hand, we have trained a high-precision classification model, so that each MRI can be segmented using the correct segmentation model with a high probability. For small target and multi-shape carpal tunnel segmentation tasks, the proposed Deep CTS achieved good results. Here are some examples in the test set shown in Fig. 9. Each sub-image in Fig. 9 contained the marker position (left) and predicted position (right) of a carpal tunnel.

Discussion

Carpal tunnel syndrome is a common condition that causes pain, numbness, tingling, and muscle dysfunction in the hand and arm [24]. The condition occurs when one of the major nerves to the hand, the median nerve, is squeezed or compressed as it travels through the wrist. In other words, this symptomatic compression neuropathy of the median nerve is characterized by the presence of increased pressure within the carpal tunnel and decreased function of the nerve at wrist level [25]. It can be caused by many different diseases and has no restriction of age, sex, ethnicity, or occupation.

Although CTS is a common disease in adults, it is often confused with diseases of the nervous system, musculoskeletal system, or vascular system, especially the high incidence of cervical spondylosis in middle-aged and elderly people, which leads to delay in the appropriate treatment [26, 27]. Early on, symptoms can often be relieved with simple measures like wearing a wrist splint or avoiding certain activities. If pressure on the median nerve continues, however, it can lead to nerve damage, worsening symptoms, and even muscle atrophy. To prevent permanent damage, surgery to take pressure off the median nerve may be recommended for some patients [21]. Therefore, early detection and diagnosis are paramount to ensure good outcomes in patients with CTS. Furthermore, studies show that early diagnosis also ensures better results of treatment [28, 29].

Despite its commonness, there is a paucity of evidence about the best approaches for assessment and to guide treatment decisions. Currently, clinical assessment is considered the gold standard in both clinical and research settings, but the final diagnoses would vary widely, depending on the doctor’s specialty and clinical experience. Nerve conduction study (NCS) is very sensitive in examining median nerve dysfunction caused by damage to the nerve; however, some studies showed that it has a false negative rate of about 10% and is also associated with a false positive rate of about 15% [30–34]. Furthermore, NCS is very uncomfortable for the patient, and on top of that, it cannot be used to distinguish between primary and secondary CTS. As a result, in addition to clinical assessment and NCS, imaging is also employed in the diagnosis of CTS. In the last decade or so, both ultrasound and MRI have been increasingly used to diagnose CTS, but not much data is available when we talk about the sensitivity and specificity of these diagnostic modalities [35–37]. In addition to that, their accuracy levels are unclear, and there is also the absence of clear diagnostic criteria, especially for MRI, and these need to be established to allow imaging to play a bigger role in the diagnosis of CTS [38–43]. These factors hinder the widespread acceptance of MR or ultrasound in the diagnosis of CTS, and efforts need to be taken in these areas in order to allow imaging to replace NCS in the diagnosis of CTS [44, 38, 45, 42, 43].

MRI has several advantages over ultrasound in being less operator-dependent, allowing clearer delineation of the carpal tunnel contents, and enabling the entire median nerve to be measured [46–48]. Besides, it requires only two standard axial sequences. And due to these features, MR imaging is being increasingly used to diagnose CTS [49, 44, 40, 45, 41, 43].

Some basic MRI changes have been reported in CTS as it could be used to study median nerve and direct visualization of carpal tunnel volumes. The general performance is an increased volume of median nerves at the level of the pisiform and hamate and an increased flexor retinaculum bowing in cross-sectional area and T2 signal intensity in median nerves [45, 50, 51]. Studies have stressed several different parameters to diagnose CTS. These parameters include cross-sectional area (CSA), flattening ratio (FR), and signal intensity (SI) of the median nerve and as well as palmar retinacular bowing (BR) [49, 44, 38, 45, 41]. Median nerve T2 hyperintensity which occurs due to vascular congestion, disruption of axoplasmic flow, increase in endoneural connective tissue, epineural and endoneural edema, as well as Wallerian degeneration is also a useful parameter [52, 53], and this hyperintensity on T2 weighted images is reversible after carpal tunnel release [44, 54, 55]. This feature allows for the use of MRI as a way to follow up and allow clinicians to assess the release of the carpal tunnel and tabulate the progression of the patients.

However, these complicated changes in MRI imaging may be difficult for a clinician or junior radiologist. According to statistics, the average diagnostic error rates range from 3 to 5%, with PET/CT leading the way, followed by MRI [56, 6]. One research that investigated the callback reports of outpatient CT and MRI showed that the callback rate for MRI was 0.114%, nine times that of CT; at the same time, the musculoskeletal system accounted for the bulk of the callback, regardless of CT or MRI, and was more than three times that of abdominal MRI [57]. Then, how about the error rate of the primary interpretation. Another survey of abdominal MRI reinterpretation in tertiary care hospitals showed that 68.9% of the preliminary reports had at least one error, which, combined with the previous findings, would be higher in the musculoskeletal system [58]. Inspired by the classification and segmentation model, we conducted a study on the application of deep learning in this area to allow for a greater understanding of the MRI characteristics of CTS and also to assist clinicians and radiologists to better diagnose and discern the severity of CTS in a clinical setting. This would allow for earlier diagnosis which would lead to better outcomes and increase patient satisfaction and better quality of life for the patients.

In our study, we selected MRI with a field strength of 3.0 T for it has a higher image resolution; however, since it is a relatively new imaging tool compared with MRI of 1.5 T, as well as the diagnosis of CTS by MRI is also more emerging, which is different from traditional diagnostic modalities, thus, the inclusion of patients was lacking, resulting in an insufficient amount of study data which has become a problem. Unlike other segmentation tasks that have a large number of samples, we only have data from 29 patients. Intuitively, combining MRI images of different depths of the same patient to build a 3D model will have a better effect than building a 2D model. However, the small sample size will cause the 3D model to not be well trained. In addition, the cost of 3D model training is very expensive. Therefore, we treat the MRI of the patient with different depths as an independent sample to establish a 2D carpal tunnel segmentation model. On the one hand, only one single slice is used when training a 2D model, which is equivalent to expanding the original data. On the other hand, compared with the 3D model, the 2D model also reduces the amount of calculation for training the model.

There are two types of image morphology in MRI, and they have their own image features and textures. The carpal tunnel, the target of segmentation, is too small for the total area of the image. In order to better learn the characters in each image category and segment the target better, we train a simplified U-Net segmentation network for each image type separately. The purpose of simplifying U-net is to avoid network overfitting so that small target carpal tunnels can be better segmented, which only accounts for 1.85% of the preprocessed image. In order to better help a clinician or junior radiologist, we need to realize end-to-end carpal tunnel segmentation, so we add the MRI classification network before the segmentation network.

High noise, slow scanning, complicated operation, difficult interpretations of imaging studies, and many sources of artifacts are the multiple difficulties in MRI applications. And 60–80% of diagnostic error in imaging is attributed to the initial misses of small lesion or unlooked-for one, which indicates the importance of dual reading and the benefit of artificial intelligence-assisted interpretation [6]. Meanwhile, the lack of radiologists and the uneven geographic distribution also further limit the widespread use of MRI [59]. Knocking on these problems is something we do all the time. The framework we proposed suggests that in medical image segmentation, even with a small amount of data and multiple image morphologies, we can segment small targets well, as long as we construct a suitable model for the problem. Not only that, the fast reading time of our system, 0.17 s of image segmentation, and results generation with high confidence level, compared to the average 77 s required by a general radiologist to read an MRI, or even the shortened 3–9 s required to read a simplified MRI, significantly increase the diagnostic efficiency of the doctor, saving as much time as adding a skilled technician [60, 61].

Conclusions

CTS is the most common peripheral nerve entrapment syndrome in the world, but there is a lack of widely recognized and accurate diagnostic methods. MRI is an emerging technique for the diagnosis and evaluation of CTS. It can show a variety of characteristics of this disease, and several diagnostic parameters have been found, but may challenge the clinicians and junior radiologists. Thus, we propose an end-to-end deep learning framework called Deep CTS to segment the carpal tunnel from the MR image, which can assist doctors to diagnose CTS better and faster. With the tailor-designed architecture, Deep CTS can segment the carpal tunnel region more correctly. The experimental results demonstrate that the performance of the proposed deep learning framework is better than the classical U-Net segmentation network for small carpal tunnel regions. Therefore, for the clinician or senior radiologist, Deep CTS is a powerful adjunct for the clarification of carpal tunnel syndrome while for junior surgeons and radiologists is useful to learn and draw lessons.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary file1 (DOCX 25 KB)^{(24.8KB, docx)}

Acknowledgements

This work was supported by Alibaba Cloud.

Abbreviations

CTS: Carpal tunnel syndrome
EMG: Electromyography
NCS: Nerve conduction study
MRI: Magnetic resonance image
CNN: Convolutional neural networks
FCN: Fully convolutional networks
FOV: Field of view
CLAHE: Contrast Limited Adaptive Histogram Equalization
2D: Two-dimensional
BN: Batch normalization
ReLU: Rectified linear unit
TP: True positive
TN: True negative
FP: False positive
FN: False negative
PA: Pixel accuracy
IoU: Intersection over union.

Author Contribution

HL designed the study; HY Z, AA, YZ D, and QJ B performed data collection; QB, XL H, and JY F analyzed the results; and MHAHA, SHAE, VGK, and ZW W drafted the manuscript. The authors have read and approved the final manuscript.

Funding

The study was funded by the National Natural Science Foundation of China (grant number 81702135), Zhejiang Provincial Natural Science Foundation (grant number LY20H060007, LS21H060001), Zhejiang Traditional Chinese Medicine Research Program (grant number 2017ZB057), and Alibaba Youth Studio Project (the grant number ZJU-032). The funding bodies had no role in the design of the study; in collection, analysis, and interpretation of data; and in drafting the manuscript.

Availability of Data and Materials

The dataset supporting the conclusions of this article is included with the article.

Declarations

Ethics Approval and Consent to Participate

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. The study protocols were approved by the Medical Ethics Committee of the First Affiliated Hospital of the College of Medicine, Zhejiang University (ethics approval number: 2021(224)).

Consent for Publication

Written informed consent was obtained from the patient for publication of clinical details and clinical images. Upon request, a copy of the consent form is available for review by the editor of this journal.

Competing Interests

The authors declare no competing interests.

Footnotes

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Contributor Information

Haiying Zhou, Email: 18867104351@163.com.

Qi Bai, Email: qi.bai@zju.edu.cn.

Xianliang Hu, Email: xlhu@zju.edu.cn.

Ahmad Alhaskawi, Email: ahmadalhaskawi@126.com.

Yanzhao Dong, Email: 1614016952@qq.com.

Zewei Wang, Email: 3190104382@zju.edu.cn.

Binjie Qi, Email: 1713045@zju.edu.cn.

Jianyong Fang, Email: fangjianyong@zuaa.zju.edu.cn.

Vishnu Goutham Kota, Email: vedamitravikram90@gmail.com.

Mohamed Hasan Abdulla Hasa Abdulla, Email: hamody.7.97@hotmail.com.

Sohaib Hasan Abdullah Ezzi, Email: m.siomn@yahoo.com.

Hui Lu, Email: huilu@zju.edu.cn.

References

1.Atroshi I, Gummesson C, Johnsson R, Ornstein E, Ranstam J, R. I. Prevalence of carpal tunnel syndrome in a general population. JAMA. 1999;282(2):153–158. doi: 10.1001/jama.282.2.153. [DOI] [PubMed] [Google Scholar]
2.JD, B. Do nerve conduction studies predict the outcome of carpal tunnel decompression? Muscle Nerve. 2001;24(7):935–940. doi: 10.1002/mus.1091. [DOI] [PubMed] [Google Scholar]
3.Lee JK, Yoon BN, Cho JW, Ryu HS, H. S. Carpal Tunnel Release Despite Normal Nerve Conduction Studies in Carpal Tunnel Syndrome Patients. Annals of Plastic Surgery. 2020;86(1):52–57. doi: 10.1097/sap.0000000000002570. [DOI] [PubMed] [Google Scholar]
4.Crnković T, Trkulja V, Bilić R, Gašpar D, K. R. Carpal tunnel and median nerve volume changes after tunnel release in patients with the carpal tunnel syndrome: a magnetic resonance imaging (MRI) study. Int Orthop. 2016;40(5):981–987. doi: 10.1007/s00264-015-3052-8. [DOI] [PubMed] [Google Scholar]
5.Xianliang Hu, Liu Zongyu, Zhou Haiying, Jianyong Fang HL. Deep HT: A deep neural network for diagnose on MR images of tumors of the hand. PLoS One. 2020;15(8):e0237606. doi: 10.1371/journal.pone.0237606. [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Itri JN, Tappouni RR, McEachern RO, Pesch AJ, Patel SH: Fundamentals of Diagnostic Error in Imaging. Radiographics : a review publication of the Radiological Society of North America, Inc, 38(6):1845–1865,2018. 10.1148/rg.2018180021 [DOI] [PubMed]
7.Li X, Lu J, Hu S, Cheng KK, De Maeseneer J, Meng Q, et al. The primary health-care system in China. Lancet (London, England) 2017;390(10112):2584–2594. doi: 10.1016/S0140-6736(17)33109-4. [DOI] [PubMed] [Google Scholar]
8.Shinjo D, Aramaki T: Geographic distribution of healthcare resources, healthcare service provision, and patient flow in Japan: a cross sectional study. Social Sci Med (1982) 75(11):1954–1963,2012. 10.1016/j.socscimed.2012.07.032 [DOI] [PubMed]
9.Krizhevsky Alex, Ilya Sutskever GEH. ImageNet Classification with Deep Convolutional Neural Networks. Communications of the ACM. 2017;60(6):84–90. doi: 10.1145/3065386. [DOI] [Google Scholar]
10.Szegedy Christian, Liu Wei, Jia Yangqing, Sermanet Pierre, Reed Scott, Anguelov Dragomir, Erhan Dumitru, Vincent Vanhoucke AR. Going deeper with convolutions. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2015;2015:1–9. doi: 10.1109/CVPR.2015.7298594. [DOI] [Google Scholar]
11.He K, Zhang X, S. R. and J. S. Deep Residual Learning for Image Recognition. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2016;2016:770–778. doi: 10.1109/CVPR.2016.90. [DOI] [Google Scholar]
12.Simonyan K, Zisserman A: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv 1409.1556,2014.
13.Long J, E. S. and T. D. Fully Convolutional Networks for Semantic Segmentation. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2015;2015:3431–3440. doi: 10.1109/CVPR.2015.7298965. [DOI] [Google Scholar]
14.Badrinarayanan V, Kendall A, Cipolla R. SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2017;39(12):2481–2495. doi: 10.1109/TPAMI.2016.2644615. [DOI] [PubMed] [Google Scholar]
15.Chen LC, Papandreou G, Kokkinos I, Murphy K, Y. AL. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2017;40(4):834–848. doi: 10.1109/tpami.2017.2699184. [DOI] [PubMed] [Google Scholar]
16.He K, Gkioxari G, Dollár P, Girshick R: Mask R-CNN. 2017 IEEE Int Conf Comput Vision (ICCV), 2017. 10.1109/ICCV.2017.322
17.Girshick R, Donahue J, Darrell T, Malik J: Rich feature hierarchies for accurate object detection and semantic segmentation. 580–587,2014
18.Girshick R: Fast r-cnn. 1440–1448,2015
19.Ren S, He K, Girshick R, Sun J: Faster R-CNN: towards real-time object detection with region proposal networks. Adv Neural Inf Process Syst 2015 [DOI] [PubMed]
20.Ronneberger O, Fischer P, Brox T: U-Net: Convolutional Networks for Biomedical Image Segmentation. Int Conf Med Image Comput Comput-Assist Intervention 234–241,2015. 10.1007/978-3-319-24574-4_28
21.Padua Luca, Coraci D, Erra C, Pazzaglia C, Paolasso I, Loreti C, et al. Carpal tunnel syndrome: clinical features, diagnosis, and management. The Lancet. Neurology. 2016;15(12):1273–1284. doi: 10.1016/S1474-4422(16)30231-9. [DOI] [PubMed] [Google Scholar]
22.Pizer SM, Amburn EP, Austin JD, Cromartie R, Geselowitz A, Greer T, et al. Adaptive histogram equalization and its variations. Computer vision, graphics, and image processing. 1987;39(3):355–368. doi: 10.1016/S0734-189X(87)80186-X. [DOI] [Google Scholar]
23.Milletari F, Navab N, Ahmadi SA: V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. 2016 Fourth Int Conf 3D Vision (3DV) 565–571,2016. 10.1109/3DV.2016.79
24.Preston DC, Shapiro BE: Median Neuropathy at the Wrist - ScienceDirect. Electromyography Neuromusc Disorders (Third Edition) 267–288,2013
25.Bland JD. Carpal tunnel syndrome. Curr Opin Neurol. 2005;18(5):581–585. doi: 10.1097/01.wco.0000173142.58068.5a. [DOI] [PubMed] [Google Scholar]
26.Gelfman R, Melton LJ, 3rd, Yawn BP, Wollan PC, Amadio PC, Stevens JC. Long-term trends in carpal tunnel syndrome. Neurology. 2009;72(1):33–41. doi: 10.1212/01.wnl.0000338533.88960.b9. [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Pourmemari MH, Heliövaara M, Viikari-Juntura E, Shiri R. Carpal tunnel release: Lifetime prevalence, annual incidence, and risk factors. Muscle Nerve. 2018;58(4):497–502. doi: 10.1002/mus.26145. [DOI] [PubMed] [Google Scholar]
28.Padua L, Padua R, Aprile I, Pasqualetti P, Tonali P. Multiperspective follow-up of untreated carpal tunnel syndrome: a multicenter study. Neurology. 2001;56(11):1459–1466. doi: 10.1212/wnl.56.11.1459. [DOI] [PubMed] [Google Scholar]
29.Verdugo RJ, Salinas RA, Castillo JL, Cea JG.: Surgical versus non-surgical treatment for carpal tunnel syndrome. The Cochrane database of systematic reviews, 2008(4):CD001552,2008. 10.1002/14651858.CD001552.pub2 [DOI] [PMC free article] [PubMed]
30.Jablecki CK, Andary MT, So YT, Wilkins DE, W. F. Literature review of the usefulness of nerve conduction studies and electromyography for the evaluation of patients with carpal tunnel syndrome. AAEM Quality Assurance Committee. Muscle Nerve. 1993;16(12):1392–1414. doi: 10.1002/mus.880161220. [DOI] [PubMed] [Google Scholar]
31.Jarvik JG, Yuen E, Kliot M. Diagnosis of carpal tunnel syndrome: Electrodiagnostic and MR imaging evaluation. Neuroimaging Clinics of North America. 2004;14(1):93–102. doi: 10.1016/j.nic.2004.02.002. [DOI] [PubMed] [Google Scholar]
32.Mackinnon SE. Pathophysiology of nerve compression. Hand Clinics. 2002;18(2):231–241. doi: 10.1016/S0749-0712(01)00012-9. [DOI] [PubMed] [Google Scholar]
33.Padua Luca, Padua R, Aprile I, D’Amico P, Tonali P. Carpal tunnel syndrome: Relationship between clinical and patient-oriented assessment. Clinical Orthopaedics and Related Research. 2002;395:128–134. doi: 10.1097/00003086-200202000-00013. [DOI] [PubMed] [Google Scholar]
34.Wright SA, L. N. Nerve conduction studies as a routine diagnostic aid in carpal tunnel syndrome. Rheumatology (Oxford) 2003;42(4):602–603. doi: 10.1093/rheumatology/keg138. [DOI] [PubMed] [Google Scholar]
35.Roll SC, Case-Smith J, Evans KD: Diagnostic accuracy of ultrasonography vs. electromyography in carpal tunnel syndrome: a systematic review of literature. Ultrasound Med Biol 37(10):1539–1553,2011. 10.1016/j.ultrasmedbio.2011.06.011 [DOI] [PubMed]
36.Brienza M, Pujia F, Colaiacomo MC, Anastasio MG, Pierelli F, Di Biasi C, et al: 3T diffusion tensor imaging and electroneurography of peripheral nerve: a morphofunctional analysis in carpal tunnel syndrome. J Neuroradiol = J de Neuroradiologie 41(2):124–130,2014. 10.1016/j.neurad.2013.06.001 [DOI] [PubMed]
37.Barcelo C, Faruch M, Lapègue F, Bayol M-A, Sans N. 3-T MRI with diffusion tensor imaging and tractography of the median nerve. European radiology. 2013;23(11):3124–3130. doi: 10.1007/s00330-013-2955-2. [DOI] [PubMed] [Google Scholar]
38.Jarvik JG, Kliot M, M. K. MR nerve imaging of the wrist and hand. Hand Clin. 2000;16(1):13–24. doi: 10.1016/S0749-0712(21)00498-4. [DOI] [PubMed] [Google Scholar]
39.Kleindienst A, Hamm B, Lanksch WR. Carpal tunnel syndrome: Staging of median nerve compression by MR imaging. Journal of Magnetic Resonance Imaging. 1998;8(5):1119–1125. doi: 10.1002/jmri.1880080518. [DOI] [PubMed] [Google Scholar]
40.Ng AWH, Griffith JF, Lee RKL, Tse WL, Wong CWY, Ho PC. Ultrasound carpal tunnel syndrome: additional criteria for diagnosis. Clinical Radiology. 2018;73(2):214.e11–214.e18. doi: 10.1016/j.crad.2017.07.025. [DOI] [PubMed] [Google Scholar]
41.Radack DM, Schweitzer ME, Taras J. Carpal Tunnel Syndrome: Are the MR Findings a Result of Population Selection Bias? American Journal of Roentgenology. 1997;169(6):1649–1653. doi: 10.2214/ajr.169.6.9393185. [DOI] [PubMed] [Google Scholar]
42.Somay G, Somay H, Çevik D, Sungur F, Berkman Z. The pressure angle of the median nerve as a new magnetic resonance imaging parameter for the evaluation of carpal tunnel. Clinical Neurology and Neurosurgery. 2009;111(1):28–33. doi: 10.1016/j.clineuro.2008.07.008. [DOI] [PubMed] [Google Scholar]
43.Tsujii M, Hirata H, Morita A, Uchida A. Palmar Bowing of the Flexor Retinaculum on Wrist MRI Correlates With Subjective Reports of Pain in Carpal Tunnel Syndrome. Journal of Magnetic Resonance Imaging. 2009;29(5):1102–1105. doi: 10.1002/jmri.21459. [DOI] [PubMed] [Google Scholar]
44.Cha JG, Han JK, Im S. Bin, Kang SJ. Median nerve T2 assessment in the wrist joints: Preliminary study in patients with carpal tunnel syndrome and healthy volunteers. Journal of Magnetic Resonance Imaging. 2014;40(4):789–795. doi: 10.1002/jmri.24448. [DOI] [PubMed] [Google Scholar]
45.Pasternack II, Malmivaara A, Tervahartiala P, Forsberg H, V. T. Magnetic resonance imaging findings in respect to carpal tunnel syndrome. Scand J Work Environ Health. 2003;29(3):189–196. doi: 10.5271/sjweh.721. [DOI] [PubMed] [Google Scholar]
46.Britz GW, Haynor DR, Kuntz C, Goodkin R, Gitter A, K. M. Carpal tunnel syndrome: correlation of magnetic resonance imaging, clinical, electrodiagnostic, and intraoperative findings. Neurosurgery. 1995;37(6):1097–1103. doi: 10.1227/00006123-199512000-00009. [DOI] [PubMed] [Google Scholar]
47.Dailey AT, Tsuruda JS, Filler AG, Maravilla KR, Goodkin R, K. M. Magnetic resonance neurography of peripheral nerve degeneration and regeneration. Lancet. 1997;350(9086):1221–1222. doi: 10.1016/S0140-6736(97)24043-2. [DOI] [PubMed] [Google Scholar]
48.Duncan I, Sullivan P, L. F. Sonography in the diagnosis of carpal tunnel syndrome. AJR Am J Roentgenol. 1999;173(3):681–684. doi: 10.2214/ajr.173.3.10470903. [DOI] [PubMed] [Google Scholar]
49.Allmann KH, Horch R, Uhl M, Gufler H, Altehoefer C, Stark GB, Langer M. MR imaging of the carpal tunnel. European Journal of Radiology. 1997;25(2):141–145. doi: 10.1016/S0720-048X(96)01038-8. [DOI] [PubMed] [Google Scholar]
50.Martins RS, Siqueira MG, Simplício H, Agapito D, Medeiros M. Magnetic resonance imaging of idiopathic carpal tunnel syndrome: Correlation with clinical findings and electrophysiological investigation. Clinical Neurology and Neurosurgery. 2008;110(1):38–45. doi: 10.1016/j.clineuro.2007.08.025. [DOI] [PubMed] [Google Scholar]
51.W, B. Radiologic imaging of the carpal tunnel. Eur J Radiol. 1997;25(2):112–117. doi: 10.1016/s0720-048x(97)00038-7. [DOI] [PubMed] [Google Scholar]
52.Horch RE, Allmann KH, Laubenberger J, Langer M, Björn Stark G. Median nerve compression can be detected by magnetic resonance imaging of the carpal tunnel. Neurosurgery. 1997;41(1):76–83. doi: 10.1097/00006123-199707000-00016. [DOI] [PubMed] [Google Scholar]
53.Teresi LM, Hovda D, Seeley AB, Nitta K, Lufkin RB. MR imaging of experimental demyelination. American Journal of Neuroradiology. 1989;10(2):307–314. doi: 10.2214/ajr.152.6.1291. [DOI] [PubMed] [Google Scholar]
54.Chammas M, Boretto J, Burmann LM, Ramos RM, Dos Santos Neto FC, S. J. Carpal tunnel syndrome - Part I (anatomy, physiology, etiology and diagnosis) Rev Bras Ortop. 2014;49(5):429–436. doi: 10.1016/j.rboe.2014.08.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
55.J., C. A coefficient of agreement for nominal scales. Educ Psych Meas. 1960;20:37–46. doi: 10.1177/001316446002000104. [DOI] [Google Scholar]
56.Brigham LR, Mansouri M, Abujudeh HH. JOURNAL CLUB: Radiology Report Addenda: A Self-Report Approach to Error Identification, Quantification, and Classification. AJR. American journal of roentgenology. 2015;205(6):1230–1239. doi: 10.2214/AJR.15.14891. [DOI] [PubMed] [Google Scholar]
57.Liles AL, Francis IR, Kalia V, Kim J, Davenport MS. Common Causes of Outpatient CT and MRI Callback Examinations: Opportunities for Improvement. AJR. American journal of roentgenology. 2020;214(3):487–492. doi: 10.2214/AJR.19.21839. [DOI] [PubMed] [Google Scholar]
58.Kostrubiak DE, DeHay PW, Ali N, D’Agostino R, Keating DP, Tam JK, Akselrod DG. Body MRI Subspecialty Reinterpretations at a Tertiary Care Center: Discrepancy Rates and Error Types. AJR. American journal of roentgenology. 2020;215(6):1384–1388. doi: 10.2214/AJR.20.22797. [DOI] [PubMed] [Google Scholar]
59.Nishie A, Kakihara D, Nojo T, Nakamura K, Kuribayashi S, Kadoya M, et al. Current radiologist workload and the shortages in Japan: how many full-time radiologists are required? Japanese journal of radiology. 2015;33(5):266–272. doi: 10.1007/s11604-015-0413-6. [DOI] [PubMed] [Google Scholar]
60.Maurer MH, Brönnimann M, Schroeder C, Ghadamgahi E, Streitparth F, Heverhagen JT, et al. Time Requirement and Feasibility of a Systematic Quality Peer Review of Reporting in Radiology. RoFo : Fortschritte auf dem Gebiete der Rontgenstrahlen und der Nuklearmedizin. 2021;193(2):160–167. doi: 10.1055/a-1178-1113. [DOI] [PubMed] [Google Scholar]
61.Renzulli M, Brocchi S, Pettinari I, Biselli M, Clemente A, Corcioni B, et al. New MRI series for kidney evaluation: Saving time and money. The British journal of radiology. 2019;92(1099):20190260. doi: 10.1259/bjr.20190260. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary file1 (DOCX 25 KB)^{(24.8KB, docx)}

Data Availability Statement

The dataset supporting the conclusions of this article is included with the article.

PERMALINK

Deep CTS: a Deep Neural Network for Identification MRI of Carpal Tunnel Syndrome

Haiying Zhou

Qi Bai

Xianliang Hu

Ahmad Alhaskawi

Yanzhao Dong

Zewei Wang

Binjie Qi

Jianyong Fang

Vishnu Goutham Kota

Mohamed Hasan Abdulla Hasa Abdulla

Sohaib Hasan Abdullah Ezzi

Hui Lu

Abstract

Supplementary Information

Background

Materials

Patient Data

Description of the Dataset

Fig. 1.

Data Preparation

Fig. 2.

Methods

Samples

Table 1.

Framework of the Data Flow

Fig. 3.

Conv Block

Fig. 4.

Classification Model

Fig. 5.

U-Net

Fig. 6.

Training and Evaluation Criteria

Experimental Results

Table 2.

Fig. 7.

Fig. 8.

Fig. 9.

Discussion

Conclusions

Supplementary Information

Acknowledgements

Abbreviations

Author Contribution

Funding

Availability of Data and Materials

Declarations

Ethics Approval and Consent to Participate

Consent for Publication

Competing Interests

Footnotes

Contributor Information

References

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases