Automatic segmentation of extensor carpi ulnaris tendon and detection of tendinosis with convolutional neural networks

Abstract

Background

Extensor Carpi Ulnaris (ECU) tendinosis, a frequent cause of chronic wrist pain, requires prompt diagnosis to prevent disability. This study demonstrates the use of convolutional neural networks (CNNs) for automated detection and segmentation of the ECU tendon and tendinosis in 2D axial wrist MRI.

Purpose

To develop a CNN for the automated detection of ECU tendon and automatic delineation of tendinosis in 2D wrist MRI. The study serves as a proof-of-concept, demonstrating the feasibility of automating the segmentation of musculoskeletal structures in wrist MRI and offering an efficient solution for detecting tendinosis.

Material and Methods

In a retrospective analysis of 1081 patients undergoing wrist MRI imaging, 46 patients exhibited tendinosis. Two deep learning-based methods for segmenting the ECU tendon and T2 hyperintense lesions indicative of tendinosis from 2D axial wrist MRI series were developed and compared in this study. Both methods were trained and evaluated over all 46 patients using Dice score as the main evaluation metric.

Results

The mean ECU tendon segmentation Dice score ranged from 0.61 to 0.64 (± 0.27 to 0.31). Tendinosis detection yielded a Dice score of 0.38 for both the threshold method (±0.19) and the CNN (±0.22). A Dice score > 0.50 indicated successful detection, with our methods achieving a detection rate of 72–76%.

Conclusion

The developed CNN effectively detected and segmented the ECU tendon in 2D MRI series. Tendinosis was detected with comparable accuracy using both signal intensity thresholding and the trained CNN method.

Keywords

segmentation MR-imaging tendons computer applications-detection diagnosis wrist

Introduction

The extensor carpi ulnaris (ECU) tendon, integral for wrist extension and ulnar deviation, is susceptible to tendinosis, stemming most commonly from repetitive use or degeneration. Proposed reasons for tendinosis include mechanical overload and microtrauma, pathological responses (Prostaglandin E2 and inflammatory cell infiltration), and aging process of tendon stem cells.^1,2

ECU tendinosis significance lies in its potential to cause chronic pain and impair wrist function. Prompt diagnosis is crucial to avoid long-term disability, especially in populations engaged in repetitive wrist motions.

Treatment strategies for ECU tendinosis involve conservative approaches, including rest, physical therapy, and corticosteroid injections. Surgical options can be entertained in persistent cases of tendinosis where co-compartment release could prove beneficial.³ In cases of concurrent subluxation or instability, sub-sheath repair or reconstruction has shown good post-operative outcomes.^1,4 MRI imaging plays a pivotal role in diagnosing and determining the treatment course, including assessing candidacy for surgical procedures.

MRI findings in tendinosis are distinctive, showing tendon thickening or thinning and T2 hyperintense lesions.⁵ These lesions are associated with disorganized collagen fibers and increased water content.⁶ While the significance of T2 hyperintense lesions is debated—appearing in both asymptomatic and symptomatic⁷ patients, and sometimes linked to peripheral TFCC tears⁸—this study focuses on T2 hyperintense lesions as a marker of tendinosis.⁹

In the realm of wrist MRI, advancements are observed in areas such as cartilage segmentation and segmentation of bone marrow edema. Additionally, there is notable presence of studies focusing on automated detection of osteoarthritis and rheumatic diseases. Artificial intelligence in tendon segmentation, identification, and detection of pathologic changes has been done before in musculoskeletal MRI, in the form of quantifying tenosynovitis¹⁰ in wrist MRI, knee ACL tear identification,¹¹ quadriceps and patellar tendon mapping,¹² Achilles’ tendon tear mapping,¹³ peroneal tendon mapping,¹⁴ shoulder rotator cuff tendon tear identification,¹⁵ and gluteal tendinopathy mapping.¹⁶

In the current literature, the segmentation of ECU tendon and possible associated tendinosis are not well represented. This study is a proof-of-concept where we trained a convolutional neuronal network (CNN) to automatically identify and segment the ECU tendon and additionally to delineate tendinosis.

Materials and methods

We conducted a retrospective study of the electronic medical records in a tertiary hospital. All primary MRI imaging studies with a unique patient (n = 1081) aged >15 years who underwent at least a finger, wrist, or hand imaging referred by a hand surgeon or resident from Töölö University Hospital in Helsinki between January 2016 and June 2017 were included in the study (Fig. 1). The patients were all over 15 years old, as the selected hospital only treated older patients; younger patients are directed to a children’s hospital.

Figure 1.

Study inclusion flow chart.

From the dataset, we identified 4% (n = 48) with prominent high T2 or PD fs signal within the tendon constituting tendinosis. The material consisted of 23 males, 23 females, with a mean age of 45.43 years (±14.58).

Orientation of study data were 76% (n = 35) with the phalanges pointing superiorly, 20% (n = 9) with phalanges pointing inferiorly, and 2% (n = 1) were oblique and 2 % (n = 1) were horizontal. Also, in 93% (n = 43), the axial series was oriented as having the palmar side pointing toward the superior edge of the image, and in 7% (n = 3), the volar side pointed inferiorly. FOVs of the studies consisted of only the wrist in 87 % (n = 40), wrist and palm in 7% (n = 3), whole hand in 4% (n = 2), and wrist and distal antebrachium in 2% (n = 1). Each study was imaged either with a GE Optima 430 1.5T or Sigma HDxt 1.5T MR scanner. Routine clinical axial, sagittal, and coronal MRI 2D sequences were used including coronal and sagittal T1, PD with fat saturation, gradient-echo T2, and axial PD and PD with fat saturation. From the studies, only axial proton density fat-saturated or T2 fat-saturated series were segmented and trained.

Manual segmentation

The axial series of each study were manually segmented using 3D Slicer 5.4.0¹⁷ by the radiologist (MH). The ECU tendon was segmented fully in the field of view, right until its insertion in the 5th metacarpal base. Tenosynovitis was noted in 11% of cases (n = 5), and in 2% (n = 1) of cases, pulsation of nearby artery caused high signal artefact in the ECU tendon, but neither finding was included in the segmentation.

Tendinosis was segmented as a separate entry. High signal areas caused by artefacts (mostly golden angle or pulsation) were delineated. An example segmentation is illustrated in Fig. 2.

Figure 2.

Example of manual segmentation. Left: Radiologist segmented axial slice of PD FS series. Right: Corresponding unsegmented slice.

Automatic segmentation

Two different methods, visualized in Fig. 3, for automatic segmentation of ECU tendon and tendinosis were experimented. In the first (1) method, 2D V-Net model was used for segmenting both the ECU tendon and tendinosis slice by slice. The architecture of the model followed the MONAI version 0.8.0 implementation of the one proposed in a previous study.¹⁸ In the architecture, 2D convolutional layers replaced their 3D counterparts, the number of convolution layers was slightly different than in the previous study, batch normalization and dropout was applied and exponential linear units were used instead of parametric linear units. The architecture is visualized in more detail in Fig. 4. The model predictions were post-processed by removing all but the largest ECU tendon prediction in terms of z-axis length. All tendinosis predictions outside the largest prediction were also removed.

Figure 3.

Two methods used in this study for segmentation of ECU tendon and tendinosis. In the first, 2D V-Net segments axial wrist MRI-scans slice by slice and produces both 3D ECU tendon (shown in green) and tendinosis segmentation predictions (shown in yellow). In the second, another 2D V-Net creates only ECU tendon prediction and tendinosis is defined based on a thresholding algorithm. The algorithm assigns voxels inside ECU tendon prediction volume as tendinosis voxels based on given intensity percentile and uses erosion operations to smooth the result.

Figure 4.

The neural network architecture of the V-Net models used in this study. Output dimension n is two with method 2 and three with method 1. Padding is applied to all convolution layers except those with 2 × 2 filters. The bias parameter is not learned in any of the convolutional layers.

6 V-Net models were trained with using a maximum of 8 studies for validation and testing and 38 for training. The combined test set of all models included all 46 samples of this study. Each model was trained with the parameters illustrated in Table 1.

Table 1.

Neural network training parameters.

Optimizer	Adam
Learning rate	0.0001, halved after 1000 and 1500 epochs
Batch size	32 with 16 samples collected from two different studies
Patch size	512 × 512 × 1
Augmentation	Rotation from 0 to 180°, applied with 0.5 probability to batch samples
Maximum number of epochs	2000
Loss function	Summation of Dice score and categorical cross entropy (binary with method 2)
Model configuration update interval	25 epochs
Model configuration update metric	Average Dice score over validation studies

In the second (2) method, V-Net model was used only to delineate ECU tendon and tendinosis segmentation was created via automatic threshold-based algorithm. The architecture of the model was the same as in the method 1, but output dimensions were 2 instead of 3. The algorithm first eroded the ECU tendon slice by slice with one pixel kernel size. Then, the ECU tendon was segmented based on a given intensity percentile threshold and the voxels above this percentile were determined as tendinosis. Tendinosis predictions of each slice with less than 3 pixels were removed as a smoothing measure. In the method, again 6 V-Net models were trained with the same training setup and otherwise the same architecture, but output channels set to two, as in the method 1. In addition, the post-processing operation of the method 2 was also used to ECU tendon predictions before applying the algorithm.

Evaluation metrics

Evaluation metrics consisted of Dice score and detection score. The detection rate was derived from Dice score and had a binary output: 1 if Dice score between predicted and reference segmentation was over 0.5 and 0 otherwise. Mean and standard deviations of these scores were calculated over all combined 46 studies of the test sets of methods 1 and 2.

Results

The mean ECU tendon segmentation Dice score was 0.61–0.64 (±0.27–0.31), depending on which of the two methods were used. Dice score >0.50 indicated successful detection, achieving a detection rate of 72%–76%. A successful detection is shown in Fig. 5, a failed prediction is shown in Fig. 6.

Figure 5.

Successful detection of the same case as in Fig. 2. Left: Successful automatic segmentation. Right: Corresponding unsegmented slice.

Figure 6.

Unsuccessful prediction. Left: base of MCV has been segmented. Right: Corresponding unsegmented slice.

In the first method, a CNN was trained to predict ECU and tendinosis, and in the second method, the ECU was predicted, and signal intensity thresholds were used to predict tendinosis. Results from both methods are presented in Table 2, with all threshold variations for the second method detailed in Table 3.

Table 2.

Regardless of the method used, the results obtained with the trained CNN and the threshold method were comparable.

Method	Tendinosis detection Dice score
1. Trained CNN method	0.38 ( ± 0.22)
2. Threshold method	0.38 ( ± 0.19)

Table 3.

In the table are shown the percentiles of signal intensity threshold values. The resulting predictions are compared to radiologist-segmented tendinosis, depicted as Dice scores. The optimal threshold for signal intensity was at the 70-75th percentiles.

Percentile	60	65	70	75	80	85	90	95
Mean tendinosis Dice score ( ± SD)	0.36 ( ± 0.18)	0.37 ( ± 0.18)	0.38 ( ± 0.19)	0.38 ( ± 0.19)	0.38 ( ± 0.2)	0.37 ( ± 0.2)	0.33 ( ± 0.2)	0.24 ( ± 0.17)

The breakdown of false positives after post-processing is as follows: In the first method, other tendons were incorrectly identified in 5 out of 6 false positives, with one instance involving the base of the 5th metacarpal (shown in Fig. 6). Additionally, a dorsal metacarpal vein was predicted in one case. In the second method, three false positives involved other tendons, while in one study, no predictions were made.

Discussion

The extensor carpi ulnaris (ECU) tendon, integral for wrist extension and ulnar deviation, can be a source of discomfort and impairment. Biophysical properties such as water content and disorganization collagen fibers⁶ of tendons define their T2 signal in MRI. Tendinosis can signify T2 signal increase relating to degeneration and tearing.⁸ This makes detecting ECU tendon abnormalities in wrist MRI studies meaningful. Here, we present a method for detecting the ECU tendon and tendinosis using a CNN to generate predictions, along with a secondary approach that employs a CNN for tendon detection and a relative signal threshold method for tendinosis detection. This is a proof-of-concept study, showing that single tendon detection and tendinosis changes can be automatically detected and segmented with reasonable accuracy.

Segmentation of various tendons in the carpal tunnel of the wrist has been done before,¹⁹ our study builds upon this by segmenting a single specific tendon and being able to delineate tendinosis. Segmentation of signal increase in tendons has been done in the supraspinatus tendon⁵, patellar tendon,²⁰ and the Achilles’ tendon,²¹ but not in the wrist—and studies aiming to do both are not well represented in current research.

In our method, the CNN successfully detected the ECU tendon along with tendinosis with respectable precision in 72%–76% of studies. We achieved a mean ECU Dice score of 0.61–0.64 for tendon segmentation and a Dice score of 0.38 for tendinosis detection compared to radiologist-segmented series using both methods. False positives primarily involved the predictions of other tendons, as well as the dorsal metacarpal vein and bony structures. False positives could be reduced by using more consistent FOVs in training data and the orientation of the wrist could be constant, but the inconsistencies in the data represent normal hospital imaging studies that could be seen as a strength of the created algorithm.

In this study, only axial 2D series were used, no coronal or sagittal series were used and no 3D series. Incorporating orthogonal series might improve the results, but because of the orientation of the ECU tendon in wrist MRI, we suggest that the axial plane is representative of the capabilities of CNN automatic segmentation. We recognize that higher field strengths (3T), thinner slices, and optimizing protocol echo times for tendon imaging could further improve the detection of the ECU tendon and associated tendinosis.

One of the limitations of this study was that there were now separate test set, as the methods 1 and 2 were tested via cross validation. However, no hyperparameter tuning for 2D V-Nets of the methods was performed, except for the choice of amount of dropout and rotation augmentation. With the rotation augmentation, we tried to address the variation of location of ECU tendons in the series images and dropout was used to prevent overfitting. 2D V-Nets used otherwise their baseline architectures freely available in MONAI. Other training parameters were chosen based on previous promising results on a different segmentation task.²² Cross-validation with separate validation and test sets allowed us to use all 46 studies as a combined test set and to ensure that no neural network weight parameters were updated based on the test set studies.

In conclusion, we demonstrated a proof-of-concept method that shows promising results in detecting the ECU tendon and tendinosis in 2D MRI axial series. This type of automatic segmentation of wrist tendinosis does not have representation in current literature, and pairing with accurate tendon segmentation offers new perspectives into the use of CNN in MRI reading that may benefit patient and clinician. This study holds clinical significance as timely detection of ECU tendinosis is important for preventing chronic pain and preserving wrist function. Drawing attention to abnormal findings in MRI imaging aids in diagnosis and avoiding long-term disability, particularly in individuals with repetitive wrist motions. Our research contributes to advancing musculoskeletal MRI techniques, offering a method for automatically detecting both tendon and tendinosis. This is of importance given the role of MRI in guiding treatment decisions.

Footnotes

Acknowledgments

The authors wish to thank the Finnish Computing Competence Infrastructure (FCCI) for supporting this project with computational and data storage resources.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Helsinki University Hospital [MK: Y780024033].

Ethical statement

ORCID iDs

Mathias Hämäläinen

Sauli Savolainen

References

Garcia-Elias

. Tendinopathies of the extensor carpi ulnaris. Handchir Mikrochir Plast Chir 2015; 47: 281–289.

Zheng

Liu

, et al. Low-intensity pulsed ultrasound promotes the repair of Achilles tendinopathy by downregulating the JAK/STAT signaling pathway in rabbits. IEEE Trans Ultrason Ferroelectrics Freq Control 2024; 71: 141–152.

Ahmad

Michael

. Stenosing tenosynovitis of the extensor carpi ulnaris. J Hand Surg 1986; 11: 519–520.

Lari

Burhamah

Alherz

, et al. Operative management of extensor carpi ulnaris instability. A Systematic Review J Wrist Surg 2023; 13: 272–281.

Pow

Bokor

Deady

, et al. Grading the severity of the rotator cuff tendinosis on MRI: assessment of inter-observer agreement and evaluation of a novel objective assessment tool. J Med Imaging Radiat Oncol 2022; 66: 357–361.

Xie

Liu

, et al. Quantitative magnetic resonance imaging-based tendon healing of different regions of the shoulder: comparison between the suture-bridge and single-row techniques. Orthop J Sports Med 2023; 11: 23259671221137835.

Kuntz

Janssen

Ring

. Incidental signal changes in the extensor carpi ulnaris on MRI. Hand (N Y) 2015; 10: 750–755.

Nevalainen

Zoga

Rivlin

, et al. Extensor carpi ulnaris tendon pathology and ulnar styloid bone marrow edema as diagnostic markers of peripheral triangular fibrocartilage complex tears on wrist MRI: a case-control study. Eur Radiol 2023; 33: 3172–3177.

Weinreb

Sheth

Apostolakos

, et al. Tendon structure, disease, and imaging. Muscles Ligaments Tendons J 2014; 4: 66–73.

10.

Aizenberg

Shamonin

Reijnierse

, et al. Automatic quantification of tenosynovitis on MRI of the wrist in patients with early arthritis: a feasibility study. Eur Radiol 2019; 29: 4477–4484.

11.

Awan

Rahim

MSM

Salim

, et al. Efficient detection of knee anterior cruciate ligament from magnetic resonance imaging using deep learning approach. Diagnostics 2021; 11: 105.

12.

Krämer

Maggioni

Brisson

, et al. T1 and T2* mapping of the human quadriceps and patellar tendons using ultra-short echo-time (UTE) imaging and bivariate relaxation parameter-based volumetric visualization. Magn Reson Imaging 2019; 63: 29–36.

13.

Regulski

Zieliński

. Multi-step segmentation algorithm for quantitative magnetic resonance imaging T2 mapping of ruptured Achilles tendons. IEEE Access 2020; 8: 199995–200004.

14.

Wilson

Surowiec

Johnson

, et al. T2* mapping of peroneal tendons using clinically relevant subregions in an asymptomatic population. Foot Ankle Int 2017; 38: 677–683.

15.

Kim

Park

Kim

, et al. Towards diagnosis of rotator cuff tears in 3-D MRI using 3-D convolutional neural networks. In: Workshop on computational biology at the international conference on machine learning, proceedings, Long Beach, CA, 14 Jun 2019, 1–7.

16.

Dadour

Gilbert

Lepage-Saucier

, et al. Quantitative MRI in patients with gluteal tendinopathy and asymptomatic volunteers: initial results on T1- and T2*-mapping diagnostic accuracy and correlation with clinical assessment. Skeletal Radiol 2021; 50: 2221–2231.

17.

Kikinis

Pieper

Vosburgh

. 3D slicer: a platform for subject-specific image analysis, visualization, and clinical support. Intraoperative Imaging Image-Guided Therapy 2014; 3: 277–289.

18.

Milletarì

Navab

Ahmadi

. V-Net: fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV), Stanford, CA, 25–28 Oct 2016, pp. 565–571, 3DV.

19.

Yang

Sun

, et al. A fully-automatic segmentation of the carpal tunnel from magnetic resonance images based on the convolutional neural network-based approach. J Med Biol Eng 2021; 41: 610–625.

20.

Kijowski

Wilson

Liu

. Bicomponent ultrashort echo time T2* analysis for assessment of patients with patellar tendinopathy. J Magn Reson Imag 2017; 46: 1441–1447.

21.

Juras

Apprich

Szomolanyi

, et al. Bi-exponential T2 analysis of healthy and diseased achilles tendons: an in vivo preliminary magnetic resonance study and correlation with clinical score. Eur Radiol 2013; 23: 2814–2822.

22.

Kesävuori

Kaseva

Salli

, et al. Deep learning-aided extraction of outer aortic surface from CT angiography scans of patients with Stanford type B aortic dissection. Eur Radiol Exp 2023; 7: 35.