Interactive Segmentation of Compressed Spinal Canal and Cord in Degenerative Cervical Myelopathy

doi:10.21203/rs.3.rs-4899028/v1

Study Design: Retrospective Diagnostic Study

Objective: We aim to develop an interactive segmentation model that can offer accuracy and reliability for the segmentation of the compressed spinal cord in degenerative cervical myelopathy (DCM).

Setting:Boramae Medical Center, Korea.

Methods: A dataset of 1,444 frames from 294 MRI records of DCM patients was used and we developed two different segmentation models for comparison: autosegmentation and interactive segmentation. The former was based on U-Net and utilized a pretrained ConvNeXT-tiny as its encoder. For the latter, we employed an interactive segmentation model structured by SimpleClick, a large model that utilizes a vision transformer as its backbone, together with simple fine-tuning. The segmentation performances of the two models were compared in terms of their DICE scores. The efficiency of the interactive segmentation model was evaluated by the number of clicks required to achieve a target mean intersection over union (mIoU).

Results: The auto and interactive segmentation models with 10 clicks returned a 0.8226 and 0.9537 DICE score for cases involving canal segmentation and a 0.7363 and 0.7767 DICE score for cases involving cord mask segmentation alone, respectively. The required clicks for the interactive segmentation model to achieve a 90% mIoU for spinal canal with cord cases and 80% mIoU for spinal cord cases were 11.71 and 11.99, respectively.

Conclusions: We found that the interactive segmentation model outperformed the autosegmentation model. Simple manual inputs can help the model identify a region of interest in the irregular shape of spinal cord.

Sponsorship: No sponsorship

Health sciences/Medical research/Preclinical research

Biological sciences/Neuroscience/Spine regulation and structure/Spine structure

Cervical

Cord

Interactive

Myelopathy

Segmentation

Degenerative cervical myelopathy (DCM) is a condition that may include cervical spondylotic myelopathy, degenerative disc disease, and ligamentous aberrations such as ossification of the posterior longitudinal ligament[1]. DCM is a progressive disease and a common cause of cervical spinal cord dysfunction, especially in elderly patients[2]. The characteristics of DCM are that the symptoms of myelopathy in 20–60% of patients with DCM will worsen over time and the deterioration is stepwise rather than steady[3, 4]. Although surgery is the main form of treatment, its primary goal is to suppress the progression of symptoms while providing only limited improvements because of the low regenerative potential of the spinal cord[5]. With an increasing elderly global population, DCM may become an important public health priority, and the importance of research may only be gradually emerging.

Magnetic resonance imaging (MRI), with its superior soft-tissue contrast capability, should provide invaluable insights into the anatomical and pathological changes within the spinal cord, facilitating early diagnosis and assessment of DCM severity[6]. The ability to visualize and quantify these changes is crucial for understanding the progression of DCM, predicting outcomes, and tailoring individual treatment plans. However, manual analysis of MRI data can be time-consuming, subjective, and prone to interobserver variability, underscoring the need for more sophisticated, reliable, and efficient analytical methods. In this context, artificial intelligence (AI), particularly advancements in computer vision, is not only emerging as a necessary and promising frontier in spinal cord research with automation of the detection, segmentation, and classification of pathological features in MRI images but also offering remarkable accuracy and speed[7]. Because image segmentation has the potential to help deep learning (DL) models improve the localization of pathological findings, diagnostic sensitivity, and thereby patient outcomes, many researchers have included segmentation stages in the construction of DL models involving spine MRI for the vertebral body and the vertebrae[8, 9]. However, in the context of DCM, the narrow spinal canal, irregularly compressed cord, and unclear borders in pathological forms can make manual segmentation of the spinal cord challenging, while autosegmentation often lacks accuracy. Therefore, some studies have introduced segmentation models for noncompressive cord lesions and detection models for cervical cord compression without segmentation[7, 10]. However, there appears to have been little research involving a spinal cord segmentation algorithm for DCM. To overcome the limitations of autosegmentation, an interactive segmentation model, which combines the precision of human expertise with the efficiency of AI-based algorithms, offers a more promising approach, given its adoption in studies involving other medical imagery[11, 12]. These methods leverage DL models, such as convolutional neural networks (CNNs), to enhance the segmentation process by providing detailed, accurate mappings of spinal cord compression from MRI scans[13]. Interactive segmentation can not only improve diagnostic reliability but also facilitate a more nuanced understanding of DCM’s impact on spinal cord anatomy, which is crucial for effective treatment planning and patient outcomes. Here, we evaluate the segmentation accuracy of an interactive image-segmentation model using simple vision transformers and compare it with other segmentation models using MRI records of patients with DCM.

This study underwent institutional review board (IRB) approval at our institution, which exempted the need for informed consent for data analysis (IRB number 30-2020-20). A waiver permission letter was obtained from the IRB administrators before data collection. Because the enrolled patients were not directly involved in this study (the data were obtained from chart reviews), informed consent was not required. Nevertheless, the extracted data from the medical records were stored confidentially. Our results were presented in accordance with the guidelines and recommendations established for AI research involving medical data [14, 15]. All MRI digital imaging and private information in medical files were anonymized before being used in the research.

Participant Selection

The MRI records and clinical data for a total of 294 patients with DCM, recorded between March 2010 and September 2022, were retrospectively reviewed. Individuals were included according to the following criteria: (a) age > 17 years, (b) had been diagnosed with mild DCM using spinal MRI, (c) had no previous history of cervical spine surgery, (d) had an exact date of diagnosis for the DCM, (e) had a baseline MRI for the same date as the diagnosis, and (f) had posterior cervical spine decompression surgery for the DCM within five years after diagnosis (74 cases) or had not (220 cases). To obtain a full range of MRI data, patients were not excluded based on (a) whether they had Parkinson’s disease or (b) the location of the DCM. More demographic details about participants are given in Table 1.

Table 1

Statistical characteristics of included patients.
Characteristics	Data
Patients (n = 294)
Female: Male, number (%)	99 (33.7): 195 (66.3)
Mean age at diagnosis ± SD, years	58.3 ± 14.0 (range, 17–90)
Number of proceeding posterior surgery	74
Primary Lesion history (n = 294)
Cervical spine surgery, number (%)	0 (0)
Different vertebrae surgery, number (%)	16 (5.4)
Cervical trauma, number (%)	67 (22.8)
Follow up duration
1st outpatient ~ Diagnosis, Mean F/U ± SD, months	8.3 ± 22.2 (range, 0–125)
1st outpatient ~ Operation, Mean F/U ± SD, months	2.5 ± 21.2 (range, 0–118)
1st outpatient ~ Last outpatient, Mean F/U ± SD, months	25.4 ± 36.7 (range, 0–150)
1st outpatient ~ Last outpatient, Mean F/U ± SD, months (without operation)	20.9 ± 36.5 (range, 0–147)
MRI findings
DCM location upper, number (%)	20 (6.8)
DCM location mid, number (%)	23 (7.8)
DCM location lower, number (%)	37 (12.6)
DCM location mix, number (%)	45 (15.3)
MCC, Mean ± SD, %	57.4 ± 13.6 (range, 17–85)
MSCC, Mean ± SD, %	6.0 ± 15.3 (range, -5.3–68.8)
Abbreviations: MRI, magnetic resonance imaging; SD, standard deviation; MSC, maximum canal compromise; MSCC, maximum spinal cord compression;

MRI Examination

Cervical spine MRI scanning was performed with a 3T (Vida, Siemens Healthineers) MRI scanner. We used images from a turbo spin echo T2-weighted sequence in the sagittal plane (T2-TSE Sag). The MRI parameter values for the T2-TSE Sag were slices for a group = 15, distance factor = 10%, position = isocenter, phase encoding direction = head to feet, phase oversampling = 50%, field of view (FOV) = 200 × 200 mm, slice thickness = 3.0 mm, repetition time (TR) = 3,500.0 ms, echo time (TE) = 82.0 ms, flip angle = 110°, average = 3, and concatenation = 1. The number of slices was occasionally different from 15 because of the variation in spine shapes in diseased patients.

Data Preprocessing and Augmentation

We utilized a labeling procedure to create a bounding box for the region of interest (RoI) and binary masks for the spinal canal and cord in 3–5 frames from each MRI recording. The selected frames in this MRI sequence were the key frame (middle frame of the sequence) and the front and back frames of the sequence. Because the spinal canal image has the cord in the middle for a sagittal-view MRI, the binary mask of the canal was represented by integrating the canal and cord area for simplicity.

Each MRI was preprocessed using a series of steps: (a) N4 bias correction to mitigate inherent biases, (b) quantile clipping to eliminate outliers among pixel values, (c) min-max normalization, (d) RoI cropping, where the RoI bounding box was labeled during the labeling phase, and (e) resizing to a uniform resolution of 256 × 256 pixels, the size mode for the cropped frames.

To ensure the stability of the pixel distribution while seeking a consistent structure for the medical images, additional augmentation methods, including rotation and random cropping, were adopted. The total dataset was partitioned into a training subset (235 cases) and a testing subset (59 cases) in an 8:2 ratio. In this way, 1,164 frames and 280 frames were utilized in the training and test phases, respectively.

Development of the Models

As shown in Fig. 1, we developed two different models to enable comparisons of segmentation performance, namely an autosegmentation model and an interactive segmentation model. Both models predict binary masks for the spinal canal and cord separately because of overlapping areas in each ground-truth mask.

For DL, an NVIDIA RTX A5000 (NVIDIA, Santa Clara, CA, USA) graphics processing unit was utilized. DL was executed using Python 3.8.10 and the PyTorch 2.0.0 framework on the Ubuntu 20.04.5 operating system. The Visual Studio Code application (Microsoft Corp., Redmond, WA, USA) was also used in the experiments.

We leveraged U-Net for the autosegmentation model, which uses ConvNeXT-tiny pretrained on ImageNet as its encoder. The randomly initialized CNN-based decoder used features encoded from each layer of the encoder as skip connections.

Recent interactive segmentation models[14, 15, 16] are large models constructed by vision transformers. Despite having many parameter settings, fine-grained supervised fine-tuning enhances the vision foundation model[17]. In this context, we employed an interactive segmentation model structured by SimpleClick[15], which uses a plain vision transformer pretrained on COCO-LVLS[18] as a backbone. The set comprising an MRI frame, the previous mask, and the number of clicks were used as the input. One reason for choosing this model was its fine-tuning performance with an MRI dataset, giving an 88.98% mean intersection over union (mIoU) result with 10 clicks on BraTS, which has a similar modality to our dataset. We evaluated both the autosegmentation and interactive segmentation models via this fine-tuning approach.

Conventional Statistical Analysis

The results for the interactive segmentation model will be modified continuously as additional manual inputs are included. Therefore, its performance could be enhanced theoretically without limit. In our experiments, we limited the number of additional clicks to 10 and evaluated it in terms of both accuracy and efficiency.

First, we obtained a DICE score to evaluate the segmentation performance in terms of accuracy. We tracked changes in the DICE score over 10 clicks to conduct a self-evaluation of the interactive segmentation model. We then compared the DICE scores for autosegmentation and interactive segmentation after the 10th click.

Next, we used the number of clicks to evaluate the models’ efficiency. We recorded the number of clicks required to achieve mIoU values of 80%, 85%, and 90%. (To avoid infeasible numbers of clicks, we limited the number to 20.)

Comparison of Performance Between the Models

The training processes for canal and cord segmentation were both performed. In each task, the models predicted binary masks, with 0 representing the background and 1 representing the RoI. Because of the complex shape of the spinal cord, the DICE scores for canal segmentation were better than for cord segmentation in both models, as shown in Table 2. The autosegmentation model obtained a DICE score of 0.8226 (± 0.0176) for the canal with cord mask segmentation and 0.7363 (± 0.0210) for the cord mask segmentation, using three different seeds. The interactive segmentation model with only one click was superior to the autosegmentation model, obtaining DICE scores of 0.8282 and 0.7516 for canal and cord mask segmentation, respectively. With additional clicks, the performance of the interactive segmentation model continued to improve, finally obtaining DICE scores of 0.9537 and 0.7767 for canal and cord mask segmentation, respectively, with 10 clicks.

Table 2

DICE Score of each segmentation model (higher scores indicating better quality). The result of the auto segmentation model is the average of trained models with three different seeds.
Model	Interactive Segmentation			Auto Segmentation
	DICE Score at			DICE Score at
Clicks	Canal	Cord	Canal		Cord
1	0.8282	0.7516	0.8226		0.7363
2	0.8712	0.7601
3	0.8900	0.7524
4	0.9052	0.7494
5	0.9183	0.7512
6	0.9295	0.7551
7	0.9384	0.7606
8	0.9438	0.7657
9	0.9500	0.7707
10	0.9537	0.7767
Mean	0.9128	0.7594	0.8226		0.7363
SD	0.0401	0.0092	0.0176		0.0210
Abbreviations: SD, standard deviation

Table 3

The number of clicks with the interactive segmentation model to achieve target mIoUs.
	# of clicks for
Area	mIoU 80%	mIoU 85%	mIoU 90%
Canal	6.28	8.14	11.71
Cord	11.99	16.32	19.58
Abbreviations: mIoU, mean intersection over union; SD, standard deviation

Efficiency of the Interactive Segmentation Model

Although the segmentation results could be improved indefinitely by continuing human guidance, the number of clicks required to achieve adequate performance is also important as a measure of efficiency. We checked the efficiency metric for achieving mIoU values of 80%, 85%, and 90% for 287 cases from the testing set, allowing a maximum of 20 clicks. The average number of clicks required to achieve an mIoU of 80% for canal and cord mask prediction was 6.28 and 11.99, respectively. Figure 1 shows a histogram of the number of cases for each number of clicks to achieve an mIoU of 80%, omitting 94 cases with 20 clicks caused by unclear or not visible border of the cord. The figure shows that 53.5% of the residual dataset cases required up to six clicks to achieve an mIoU of 80%.

DCM is a progressive disease, and the primary goal of surgery is to suppress the progression of symptoms while providing only limited improvements because of the low regenerative potential of the spinal cord[2, 5].

It is therefore important to detect the disease early, and we analyzed only mild DCM cases in this study. MRI is crucial for understanding the progression of DCM because of its ability to enable visualization and quantify temporal changes in the spinal cord. However, manual analysis of MRI data can be time-consuming, subjective, and prone to interobserver variability.

To tackle this problem, we have explored the applicability of AI for the image segmentation task in MRI analysis, which localizes the spinal canal and cord and has the potential to improve pathological findings, diagnostic sensitivity, and patient outcomes. To this end, we have developed a DL-based interactive segmentation model that should aid MRI analysis in terms of both accuracy and efficiency.

An autosegmentation model (U-Net with ConvNeXt-tiny backbone) inferred the spinal canal and cord segmentation masks of 287 patients within approximately three seconds, using our architectures, achieving average DICE scores of 0.8226 and 0.7363, respectively. This model was efficient and obtained the prediction results shown in Fig. 2. However, the interactive segmentation model using only one click outperformed the autosegmentation model for both spinal canal and cord segmentation with average DICE scores of 0.8282 and 0.7516, respectively. It was even more superior to the autosegmentation model when using 10 clicks, obtaining average DICE scores of 0.9537 and 0.7767 for spinal canal and cord segmentation, respectively. We developed the interactive segmentation tool using the source from SimpleClick[15].

To the best of our knowledge, this is the first study that has applied a DL-based interactive segmentation model to localize the compressed, irregularly shaped spinal cord in MRI images when aiming to detect the progression of DCM. We found that the interactive segmentation model outperformed the autosegmentation model when trained using 220 patients. These results demonstrate that an AI-based model with modest additional human feedback can extract successfully the complex shapes of the spinal cord from MRI data. That is, a combination of a DL model and simple manual guidance can obtain the required results both efficiently and accurately.

This study has some limitations. First, we did not involve external validation, which would be essential for verifying the broader applicability of our model. Second, we did not evaluate severe DCM cases where the spinal cord would have a more irregular or narrower border. Our plans for future work include using an external test dataset to assess the model’s performance across a variety of patient groups.

In conclusion, our results suggest that using simple manual inputs enables the DL model to better identify the RoI in the image, particularly for cases where drawing the mask is challenging, such as for the spinal cord of patients with DCM. In addition, the masks obtained via our model could be leveraged to investigate changes in the spinal cord over time.

Acknowledgements

Author contributions

Study protocol design: W.C., S.B.P.; Study supervision: W.C., S.B.P.; Data collection: S.B.P; Algorithm implementation: S.H., W.C.; Experiment conduct: S.H., W.C.; Result analysis: W.C.; Manuscript writing: C.H.K, W.C., S.B.P.; Manuscript reviewing: W.C., S.B.P.

Funding

This work was supported by the New Faculty Startup Fund from Seoul National University (Number 800-20220279).

Data Availability Statement

The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request.

Competing Interests: The authors declare no competing interests.

Hejrati N, Moghaddamjou A, Marathe N, Fehlings MG. Degenerative Cervical Myelopathy: Towards a Personalized Approach. Can J Neurol Sci 2022;49:729-40.
Takamiya S, Iwasaki M, Yokohama T, Oura D, Niiya Y, Fujimura M. The Prediction of Neurological Prognosis for Cervical Spondylotic Myelopathy Using Diffusion Tensor Imaging. Neurospine 2023;20:248-54.
Zileli M, Maheshwari S, Kale SS, Garg K, Menon SK, Parthiban J. Outcome Measures and Variables Affecting Prognosis of Cervical Spondylotic Myelopathy: WFNS Spine Committee Recommendations. Neurospine 2019;16:435-47.
Karadimas SK, Erwin WM, Ely CG, Dettori JR, Fehlings MG. Pathophysiology and natural history of cervical spondylotic myelopathy. Spine (Phila Pa 1976) 2013;38:S21-36.
Davies BM, Khan DZ, Mowforth OD, et al. RE-CODE DCM (REsearch Objectives and Common Data Elements for Degenerative Cervical Myelopathy): A Consensus Process to Improve Research Efficiency in DCM, Through Establishment of a Standardized Dataset for Clinical Research and the Definition of the Research Priorities. Global Spine J 2019;9:65S-76S.
Hilton B, Tempest-Mitchell J, Davies BM, et al. Cord compression defined by MRI is the driving factor behind the decision to operate in Degenerative Cervical Myelopathy despite poor correlation with disease severity. PLoS One 2019;14:e0226020.
Martin-Noguerol T, Onate Miranda M, Amrhein TJ, et al. The role of Artificial intelligence in the assessment of the spine and spinal cord. Eur J Radiol 2023;161:110726.
Lewandrowsk IK, Muraleedharan N, Eddy SA, et al. Feasibility of Deep Learning Algorithms for Reporting in Routine Spine Magnetic Resonance Imaging. Int J Spine Surg 2020;14:S86-S97.
Lewandrowski KU, Muraleedharan N, Eddy SA, et al. Reliability Analysis of Deep Learning Algorithms for Reporting of Routine Lumbar MRI Scans. Int J Spine Surg 2020;14:S98-S107.
Merali Z, Wang JZ, Badhiwala JH, Witiw CD, Wilson JR, Fehlings MG. A deep learning model for detection of cervical spinal cord compression in MRI scans. Sci Rep 2021;11:10473.
Feng R, Zheng X, Gao T, et al. Interactive Few-Shot Learning: Limited Supervision, Better Medical Image Segmentation. IEEE Trans Med Imaging 2021;40:2575-88.
Zhao X, Pan H, Bai W, et al. Interactive segmentation of medical images using deep learning. Phys Med Biol 2024;69.
Luo X, Wang G, Song T, et al. MIDeepSeg: Minimally interactive segmentation of unseen objects from medical images using deep learning. Med Image Anal 2021;72:102102.
Sofiiuk K, Petrov IA, Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In ICIP, 2022.
Liu X, Xu Z, Bertasius G, et al. Simpleclick: Interactive image segmentation with simple vision transformers. In ICCV, 2023.
Alexander Kirillov, Eric Mintun, Nikhila Ravi, et al. Segment Anything. In ICCV, 2023.
Xiaohu Jiang, Yixiao Ge, Yuying Ge, et al. Supervised Fine-tuning in turn Improves Visual Foundation Models. arXiv preprint arXiv:2401.10222, 2024.
Gupta, Agrim, Piotr Dollar, and Ross Girshick. "Lvis: A dataset for large vocabulary instance segmentation." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.

There is no duality of interest

Interactive Segmentation of Compressed Spinal Canal and Cord in Degenerative Cervical Myelopathy

Status:

Version 1

Abstract

Figures

Introduction

Material and Methods

Participant Selection

MRI Examination

Data Preprocessing and Augmentation

Development of the Models

Conventional Statistical Analysis

Results

Comparison of Performance Between the Models

Efficiency of the Interactive Segmentation Model

Discussion

Declarations

References

Additional Declarations

Status:

Version 1