INTRODUCTION
Degenerative cervical myelopathy (DCM) encompasses conditions like cervical spondylotic myelopathy and degenerative disc disease, often linked to ligamentous abnormalities such as ossification of the posterior longitudinal ligament [5]. DCM is a leading cause of spinal cord dysfunction in the elderly, with a progressive course where symptoms deteriorate in 20−60% of patients [9,19,23]. Surgery is typically aimed at halting symptom progression, given the spinal cord’s limited regenerative potential [1]. As the elderly population increases, the prevalence of DCM is expected to rise, making it an important public health issue.
Magnetic resonance imaging (MRI) is vital for diagnosing DCM due to its superior ability to capture soft-tissue contrasts and assess spinal cord changes [6]. Early detection and accurate assessment of DCM severity are crucial for predicting outcomes and tailoring treatments. However, manual analysis of MRI data is labor-intensive and subjective, often leading to variability between observers. This has highlighted the need for more efficient, reliable analytical methods.
Artificial intelligence (AI) and deep learning (DL), particularly in the realm of computer vision, offer significant advancements in automating the detection, segmentation, and classification of pathological features in MRI images [15]. Image segmentation plays a critical role in improving diagnostic sensitivity and patient outcomes by enabling models to better localize pathological areas. While segmentation has been incorporated into many DL models for spine MRI, its application to DCM—where the compressed spinal cord and irregular anatomy pose challenges—remains underexplored [3,8,11,12,16].
In this study, we develop and evaluate segmentation models tailored to MRI images of DCM patients using a relatively small dataset of 294 patients. We compared two types of pretrained segmentation models : an auto-segmentation model based on convolutional neural networks (CNNs) using a U-Net architecture, and an interactive segmentation model that adopts a vision transformer backbone, following the recent trend of employing ViT-based architectures in interactive segmentation [10,13,14,18,22].
We demonstrated that the performance of the interactive segmentation model—augmented by minimal manual input from human experts—surpasses that of the auto-segmentation model. In addition, we investigated the relationship between the number of manual interactions and model performance, confirming the efficiency and practicality of the interactive approach.
MATERIALS AND METHODS
This study was approved by the Institutional Review Board (IRB) of Boramae Medical Center (IRB number 30-2020-20), which waived the requirement for informed consent for data analysis. 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 have been presented in accordance with the guidelines and recommendations established for AI research involving medical data [13,18]. All MRI digital imaging and private information in medical files were anonymized before being used in the research.
Participant selection
The MRI images 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 DCM using spinal MRI, (c) had no previous history of cervical spine surgery. More demographic details about participants are given in Table 1.
MR data acquisition
Cervical spine MRI scanning was performed with a 3T (MAGNETOM Vida; Siemens Healthineers, Erlangen, Germany) 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=200×200 mm, slice thickness=3.0 mm, repetition time=3500.0 ms, echo time=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 are 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 is represented by integrating the canal and cord area for simplicity.
As shown in Fig. 1, 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, 1164 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 auto-segmentation model and an interactive segmentation model. Both models predicted binary masks for the spinal canal and cord separately because of overlapping areas in each ground-truth mask.
For more details, we leveraged U-Net for the auto-segmentation model, which uses ConvNeXT-tiny pretrained on ImageNet as its encoder. The randomly initialized CNN-based decoder was trained to predict the segmentation mask using features encoded from each layer of the encoder as skip connections.
We employed an interactive segmentation model based on SimpleClick [13], which utilizes a plain vision transformer pretrained on the COCO-LVLS [4] dataset as its backbone. The model constructed a multi-scale feature pyramid from the transformer and used it as input for a segmentation head composed of multi-layer perceptrons. The head produced a 1/4-scale probability map, which was then upsampled to generate the final output. Positive and negative user clicks were encoded into a two-channel disk map and embedded as patches. These were combined element-wise and fed into a self-attention block. The model received as input an MRI frame, the previous mask, and the number of clicks. The reason of utilizing SimpleClick for our model stems from its demonstrated performance in medical imaging tasks. For instance, in the BraTS dataset, which shared imaging modality characteristics with our DCM MRI data, SimpleClick achieved a mean intersection over union (mIoU) of 88.98% with just 10 clicks.
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.
Segmentation assessment
First, we obtained a Dice score, mIoU, Average Precision and Hausdorff distance to evaluate the segmentation accuracy. We tracked changes in the four metrics over 15 clicks to conduct a self-evaluation of the interactive segmentation model. We then compared the scores for auto-segmentation and interactive segmentation after the 15th 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).
We conducted additional statistical analyses to assess whether the observed improvements between the auto-segmentation and interactive segmentation models were statistically significant. Specifically, we applied Wilcoxon signed-rank tests to compare the performance metrics across matched test samples.
RESULTS
Comparison of performance between the models
The training processes for canal and cord segmentation are both performed. In each task, the models predicted binary masks, with 0 representing the background and one representing the RoI. The DICE scores, mIoU, Average Precision, and Hausdorff distance for canal segmentation were better than for cord segmentation in both models, which may be attributed to the complex shape of the spinal cord, as shown in Table 2.
The auto-segmentation 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 two clicks was superior to the auto-segmentation model, obtaining DICE scores of 0.8417 (+2.3%) and 0.8063 (+2.1%) 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 0.8756 (+6.4%) and 0.8228 (+11.7%) for canal and cord mask segmentation, respectively, with 15 clicks.
Our model also achieved improved segmentation accuracy in other metrics. For mIoU, our model showed 0.8896 (+1.8%) and 0.8164 (+6.0%) scores. Average Precision of our model showed 0.9589 (+3.7%) and 0.9107 (+18.2%) scores. In the Hausdorff distance metric, which is a surface distance-based metric in contrast to former metrics, our model achieved 4.36 (-53.0%) and 5.92 (-70.9%) distance values. All of these results are with 15 clicks.
Additionally, we perform Wilcoxon signed-rank tests to evaluate the statistical significance of performance differences between the auto segmentation and interactive segmentation models across different segmentation categories and click settings.
As shown in Table 3, the results demonstrate that the improvements achieved by interactive segmentation are statistically significant in all cases (p<1e-20), confirming the robustness and reliability of the observed gains.
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 (Table 4). Fig. 2 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%.
DISCUSSION
DCM is a progressive spinal cord disease, and the primary goal of surgical treatment is to halt symptom progression, as functional recovery is often limited due to the spinal cord’s low regenerative potential [1,19].
MRI is highlighted 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 to support MRI analysis in terms of both accuracy and efficiency.
Although the segmentation of the spinal cord alone may appear sufficient in some clinical settings, we intentionally segmented both the spinal cord and the spinal canal in this study. This was based on the hypothesis that the morphological relationship between these two structures—including the degree and shape of compression—may serve as important imaging features in future predictive models for neurological outcomes in DCM. An auto-segmentation model based on a U-Net architecture with a ConvNeXt-tiny backbone segmented the spinal canal and cord in 287 patients within approximately three seconds, achieving average DICE scores of 0.8226 and 0.7363, respectively. While the automated model demonstrated good efficiency shown in Fig. 3, it was outperformed by our interactive segmentation model as shown in Fig. 4 with a few user clicks. With 15 clicks, the interactive model exhibited a substantial improvement, reaching 0.8756 and 0.8228 in Dice score, respectively. These findings highlight the practical value of incorporating user input into the segmentation process. Even minimal human interaction can provide contextual cues that are difficult for fully automated models to infer—especially in cases with ambiguous boundaries or irregular morphologies, which are common in DCM. The interactive segmentation model developed here is intended to lay the foundation for such prognostic research by providing accurate and reproducible delineation of both anatomical regions. And, this observation aligns with recent trends in medical imaging, where human-in-the-loop approaches have shown promise in improving both accuracy and interpretability [20,21].
While prior studies have addressed spinal cord segmentation, many have faced limitations such as morphological variability, dependence on large annotated datasets, or lack of adaptability to irregular anatomical structures typical of DCM. Recent efforts have begun to explore more robust and efficient methods, including transformer-based and interactive approaches, which informed the development of our proposed mode [2,7,17]. To the best of our knowledge, this study is the first to apply a DL-based interactive segmentation model to localize the compressed and irregularly shaped spinal cord in MRI scans for DCM assessment. Our model, fine-tuned with a relatively small dataset of 220 patients, demonstrates that expert-guided interactions can effectively compensate for the limitations of purely automated segmentation. The interactive segmentation model highlights the accessibility and adaptability of such approaches for integration into clinical workflows.
While this study explores the applicability of both segmentation models to MRI images of DCM patients, it has some limitations. First, external validation is not performed for evaluating robustness due to challenges in obtaining diverse datasets from different institutions or imaging protocols. Second, our analysis focused exclusively on mild DCM cases; severe cases exhibiting more pronounced cord deformation or ambiguous boundaries were not included. We plan to evaluate the models on an external dataset to further assess their robustness and to extend the analysis to severe DCM cases with more complex anatomical variations.
In addition, future studies could explore using the segmentation results obtained through the interactive model to fine-tune the auto-segmentation model. This approach could enhance the auto-segmentation model’s ability to handle complex anatomical structures by leveraging high-accuracy segmentation masks as training data. Another potential avenue for improvement is to retrain the interactive segmentation model using accumulated click-based data, which may reduce the number of required clicks while improving segmentation accuracy. These adaptations could further increase the efficiency and applicability of our model in various clinical settings.
CONCLUSION
In this study, we have demonstrated 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.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation 
Print 
