Prior work shows CNN-based segmentation is effective for vertebral bodies in radiographs and MRI (Chen Y. et al., 2024; Wang H. et al., 2025). However, no public dataset provides pixel-level masks of the cervical spinous processes, so we curated a task-specific corpus of lateral cervical spine X-rays and their corresponding binary masks.

We randomly sampled 500 PNG images from the Cervical Spine X-ray Atlas (CSXA) (Ran et al., 2024) (4,963 PNGs with JSON annotations; one image per patient) and manually annotated the spinous processes on a tablet with a stylus. Annotations were converted to binary masks using an OpenCV color-segmentation pipeline: images were converted to HSV; red hue bands (0 °–10 °, 160 °–180 °) corresponding to tracings were thresholded; masks from both bands were combined; contours were extracted and filled to produce clean binary ground-truth masks. All 500 image–mask pairs were created with consistent dimensions. Trained annotators labeled the images; an expert spine surgeon (Dr. Sangwook T. Yoon) provided final review and approval. The annotation workflow is shown in Figure 2.

To ensure fair evaluation and minimize overfitting (Muraina, 2022), we used an 80/20 split, allocating 400 images for training and 100 for testing (Bichri et al., 2024); the test set remained untouched throughout all experiments (Lones, 2021). For the training images, we generated three distinct dataset variants, each used to train a separate model configuration:

Original dataset: the 400 unaltered radiographs and their corresponding masks.

All three training variants (original, CLAHE, and augmented) were derived from the same underlying set of 400 cervical radiographs. These variants reflect different preprocessing pipelines applied to identical base images rather than independent datasets, thus enabling us to evaluate how raw, intensity-normalized, and synthetically diversified training distributions affected model performance.

We trained and evaluated each model on the three dataset variants: Original (400 images); CLAHE (400 images); Augmented (800 images). The test set (n = 100) remained constant across experiments. We compared GPU vs. CPU training/inference for efficiency. All experiments used a single system with PyTorch and an NVIDIA H100 Tensor Core GPU; the uniform environment ensured consistent timing and memory profiles. The H100’s tensor cores and memory bandwidth supported large-batch training and stable convergence.

A test across five different images from the testing set show that GPU inference averaged 5–8 s/image; CPU inference on an Intel® Xeon® Gold 6,136 (2 × 12 cores) averaged 9–10 s/image. The proposed CervSpineNet is computationally lightweight (∼345 MB) and typically utilizes ∼8–9 GPU cores during inference, enabling deployment without specialized hardware.

Figure 3 shows the overall architecture and I/O. The encoder is transformer-based; the decoder is a lightweight U-Net–style module with residual and squeeze-and-excitation (SE) blocks and bilinear up-sampling.

To avoid model-class bias and follow current benchmarking guidance (Antonelli et al., 2022; Tejani et al., 2024), we compared CervSpineNet to strong and architecturally diverse baselines (Wolfrath et al., 2024): DeepLabV3+, U-Net, the full Segment Anything Model (SAM), and a text-guided SegFormer. All baselines and CervSpineNet were trained and evaluated on the same training/test split with identical pre-processing and data augmentations. U-Net, DeepLabV3+, and text-guided SegFormer were trained on images resized to 512 × 512, which is standard for CNN-based medical segmentation and keeps GPU memory requirements manageable. The SAM ViT-B and the proposed CervSpineNet were trained on 1,024 × 1,024 inputs to match the native input resolution of the SAM image encoder. All predictions were resampled back to the original image resolution for evaluation. The batch size for every training experiment was kept 1, and the models were trained on 50 epochs, selecting the checkpoint with the best testing metrics as the metrics saturated around 40–45 for all experiments. This ensured that the performance differences mainly reflect architectural choices rather than differences in data or optimization.

Following common practice (Li et al., 2025; Zhang et al., 2025), we report Dice, Intersection-over-Union (IoU), Structural Similarity (SSIM), Hausdorff Distance (HD95), and Volumetric Similarity (VS) on the held-out test set. Equations 15-17 display the Dice, IoU and VS terms as implemented in this study. D i c e = 2 · T P 2 · T P + F P + F N (15) I o U = T P T P + F P + F N (16) V S = 1 − T P + F P − T P + F N T P + F P + T P + F N (17)

Here, TP/FP/TN/FN are computed from thresholded (0.5) binary predictions and ground-truth masks for each pixel.

True Positive (TP) – prediction = 1, ground truth = 1 (pixels correctly classified as foreground)

SSIM captures luminance, contrast, and structure similarity. We used the standard SSIM metric (Bakurov et al., 2022), defined for a predicted mask x and ground truth mask y as shown in Equation 18: S S I M x , y = 2 μ x μ y + C 1 2 σ x y + C 2 μ x 2 + μ y 2 + C 1 σ x 2 + σ y 2 + C 2 (18) where, μ x   a n d   μ y are the mean pixel intensities of x and y , σ x 2   a n d   σ y 2 are the corresponding variances, σ x y is the cross-covariance between x and y ; C 1 , C 2 are small constants for stability.

HD95, as given by Equation 20, is the 95th percentile surface-to-surface distance between predicted and ground-truth objects (Ramachandran et al., 2023). Let A , B be sets of foreground boundary points with Euclidean distance d. The directed surface distance sets are calculated as in Equation 19: D A , B = min b ∈ B   d a , b | a ∈ A , D B , A = min a ∈ A   d b , a | b ∈ B (19)

Thus, for all distances, S = D A , B ∪ D B , A . H D 95 A , B = p e r c e n t i l e 95 S (20)

Ablation experiments are essential in medical image segmentation, particularly for hybrid architectures, as they clarify the relative contributions of individual components to overall performance (Gao et al., 2021; Wang X. et al., 2025). To systematically characterize the behavior of CervSpineNet, we performed two sets of ablations: (i) architectural ablations isolating encoder and decoder contributions, and (ii) loss-function ablations assessing how each loss term influences error modes and structural fidelity.

Mean performance metrics (Dice, IoU, SSIM, HD95, VS) on the testing data were computed for all models across the three dataset variants: the original set (Table 1), CLAHE-enhanced images (Table 2), and the augmented dataset (Table 3).

Repeated trials with identical hyperparameters produced standard deviations in the range 0.001–0.005 for Dice, IoU, SSIM, and VS and ∼0.3–0.6 for HD95, confirming the stability and reproducibility of model performance across the testing data. Figure 4 shows the distribution of Dice coefficients for all models evaluated on the 100 test images under the three preprocessing conditions: (A) original data, (B) CLAHE-enhanced data, and (C) augmented data. Across all three dataset variants, CervSpineNet exhibited the narrowest spread and highest median Dice scores, indicating robustness to preprocessing variations and the radiographic variability that these variants are designed to reflect.

CervSpineNet achieved the highest or near-highest mean values for nearly every metric, with particularly strong performance in SSIM and HD95, which reflect structural accuracy and boundary precision, respectively. To determine whether these improvements were statistically significant, formal non-parametric tests were conducted as described in the following section.

A summary of the mean test-set performance for all four architectural configurations adopted in this study is shown in Table 4.

Finally, the results of the loss ablation experiment are demonstrated in Table 5. Different sets of loss functions used, and their metrics scores are depicted in this table. The experiments demonstrated in Tables 4, 5 also yielded minute variations in the range of 0.002–0.004 for Dice, IoU, SSIM, and VS and variations of ∼0.45–0.70 for HD95.

Prior studies highlight the robustness of the Wilcoxon Signed-Rank test and the Friedman test, which detects overall performance differences across multiple models evaluated on the same samples, for evaluating segmentation algorithms (Mostafa et al., 2024). We compared five segmentation models (SAM, U-Net, text-guided SegFormer, DeepLabV3+, and CervSpineNet) using five metrics (Dice, IoU, SSIM, HD95, and VS) across the same per-image testing scores.

The Friedman test assessed overall model differences for each metric, followed by pairwise Wilcoxon signed-rank tests with Bonferroni correction (p < 0.05). Across all metrics, significant overall differences were observed (Friedman p < 0.001).

CervSpineNet significantly outperformed SAM and U-Net (corrected p < 0.001) in every metric and dataset variant. It also exceeded DeepLabV3+ in SSIM, HD95, and Volumetric Similarity while maintaining comparable Dice and IoU (corrected p ≈ 0.05). Against the text-guided SegFormer, CervSpineNet achieved statistically higher performance in all metrics except VS, where the difference was not significant. These results confirm that the proposed hybrid model, CervSpineNet, provides robust and consistent segmentation.

Visual inspection of unseen test radiographs further corroborated the quantitative findings (Figure 5). Panels (A), (B), and (C) show representative segmentation results from the original, CLAHE-enhanced, and augmented datasets, respectively, comparing CervSpineNet with all baseline models. Across all preprocessing conditions, CervSpineNet produced the most accurate and anatomically consistent delineations, with smoother contours, sharper boundaries, and markedly fewer false or incomplete predictions than either CNN-only or transformer-only architectures. These qualitative outcomes parallel the quantitative improvements reported in Section 3.1, illustrating how the hybrid architecture captures global contextual features while preserving fine boundary details.

To examine segmentation fidelity in more detail, Figure 6 visualizes per-pixel discrepancies between CervSpineNet predictions and expert ground-truth masks. In these heatmaps, uncolored regions indicate perfect agreement, while increasing color intensity reflects greater deviation from the annotated boundaries. False positives represent regions erroneously predicted as foreground outside the annotated structure, false negatives correspond to missed spinous process pixels within the ground-truth region, and thin color bands along the contours indicate minor boundary mismatches or thickness differences.

A complementary quantitative error assessment was performed using Dice, HD95, and Mean Absolute Error (MAE) calculated on continuous (non-binarized) probability maps to preserve boundary sensitivity. Figure 6 demonstrate a sample of radiographs with good model predictions backed by the metrics used for comparison. On the other hand, Figure 7 depicts some edge cases where the model failed to perform well due to a variety of reasons often resulting in over/under prediction of structures or significant disagreement from ground truth masks.

A pixel-by-pixel error heatmap between the ground-truth spinous process mask and the model prediction is superimposed on the lateral cervical X-ray in Figures 6, 7. Areas where prediction and ground truth agree appear like the original image, whereas colored bands around the spinous processes mark disagreement: cooler colors indicate small or no error, and warmer colors (green–yellow–red) highlight increasing mismatch, with the brightest regions corresponding to the largest segmentation errors. In these figures, the Global MAE captures the average pixel-wise deviation across the entire image, reflecting overall segmentation fidelity, whereas the Foreground-only MAE quantifies deviation exclusively within annotated spinous process regions, emphasizing local structural precision. Both metrics were computed at the native image resolution, where lower MAE values indicate stronger correspondence with expert annotations. CervSpineNet consistently achieved the lowest Global and Foreground-only MAE scores, confirming high boundary accuracy and strong structural alignment across the test set.

Figure 6 shows three representative test radiographs with high-quality CervSpineNet segmentations. In these typical cases, spinous process masks from C1–C7 closely follow the cortical margins with minimal over- or under-segmentation. These examples support the good overall performance observed in the aggregate metrics by achieving high quantitative agreement with the reference annotations (Dice ≈ 0.93, HD95 ≈ 3, Global MAE ≈ 0.005, and Foreground-only MAE ≈ 0.05).

Figure 7 illustrates three characteristic failure modes encountered by CervSpineNet. In Case 1, adjacent spinous processes are extremely close and partially overlapping, causing the model to merge two levels (C3–C4) into a single elongated mask and consequently overestimate the superior spinous process. In Case 2, a fractured spinous process produces an irregular and discontinuous contour. While the expert ground truth precisely outlines the fracture margins, the model predicts a smoother, unbroken process, resulting in reduced Dice scores and a prominent error band in the heatmap. In Case 3, the C7 spinous process is scarcely visible due to low contrast and soft-tissue overlap at the cervicothoracic junction; although the model accurately segments C1–C6, it fails to detect C7 entirely, yielding a level-specific false negative. Together, these examples indicate that severe pathology, low bone–soft-tissue contrast, and tightly packed vertebral levels are the conditions under which CervSpineNet is most vulnerable. Future work will prioritize augmenting the training set with such challenging cases and exploring level-aware constraints or post-processing strategies to mitigate mask merging and missed levels.

Figure 8 presents attention visualizations for three representative held-out test cases—two typical high-quality segmentations and one challenging example with substantial shoulder overlap. In the well-segmented cases, both attention-rollout and Grad-CAM overlays highlight a continuous chain of vertebral bodies and spinous processes spanning the upper through lower cervical levels. This pattern indicates that the transformer encoder integrates information across the full cervical column rather than relying solely on local image patches. Similarly, the query-centric attention map, when the query token is placed on a mid-cervical spinous process, assigns high attention weights to both adjacent and more distant vertebral levels, reinforcing evidence of long-range contextual reasoning. In contrast, the difficult case shows more spatially diffuse attention that partially shifts toward the overlapping shoulder and soft-tissue regions, corresponding to the observed segmentation errors. Collectively, these qualitative results support the conclusion that CervSpineNet’s transformer encoder leverages global anatomical context along the cervical spine.

We conducted a systematic analysis to compare the time required for manual versus automated segmentation of cervical spinous processes. For manual annotation, three annotators—two trained annotators and one expert spine surgeon—outlined the spinous processes on a predefined set of radiographs using a tablet interface. The total manual annotation time was determined by averaging the duration recorded across all three annotators. This process was divided into two stages: (1) annotation, measuring the average time required to trace the spinous processes, and (2) mask generation, recording the time needed to transform these annotations into binary masks using the color-based segmentation algorithm described in Section 2.1.

Figure 9 compares the time required for manual versus automated segmentation. Figure 9A summarizes the average annotation time per image for all manual annotators— two annotators and 1 expert—alongside the automated CervSpineNet model. Manual annotation times ranged from 46 to 98 s, with an overall average of 72 s, whereas CervSpineNet produced segmentations in only ∼7.5 s per image. Figure 9B shows total processing time, combining annotation and mask-generation stages. Manual processing required approximately 3–4 min per image (mean = 210 s). In contrast, automated segmentation using CervSpineNet took only 5–10 s (mean = 7.5 s), corresponding to an overall ∼96.43% reduction in total processing time.

Moreover, for inter-rater variability and to compare the model performance to human variability, we selected three of the test set images that the expert and trained annotators had already annotated and compared those segmentations to the masks predicted by the hybrid model using standard overlap and boundary metrics: Dice coefficient, Intersection-over-Union (IoU), 95th-percentile Hausdorff distance (HD95, in pixels), structural similarity index (SSIM), and volumetric similarity (VS).

Human–human agreement was high: expert vs. human 1 and expert vs. human two reached mean Dice scores of 0.91 (IoU ≈ 0.83–0.84, HD95 ≈ 6.3 pixels, SSIM ≈ 0.97, VS ≈ 0.96–0.98), while human 1 vs. human 2 yielded a similar Dice of 0.92 (IoU 0.85, HD95 6.24 pixels, SSIM 0.97, VS 0.98). On the same images, the proposed model reached a Dice of 0.92 vs. the expert (0.924), with higher IoU (0.86), lower HD95 (5.71 pixels), SSIM of 0.98, and VS of 0.99. Model vs. R1 and model vs. R2 comparisons were also in the same range: Dice ≈ 0.92, IoU ≈ 0.86, HD95 ≈ 6.8 and 5.5 pixels, SSIM ≈ 0.98. These results collectively indicate that model segmentations are as consistent with the expert as those of the additional human raters, and considering boundary accuracy, HD95 is slightly better than the average human–human variability on this sample, though we acknowledge this was based on three test cases.

This study introduces an automated segmentation framework for delineating cervical spinous processes from lateral spine radiographs. Because no public dataset existed for this task and manual annotation is labor-intensive, both a curated dataset and a novel hybrid model, CervSpineNet, were developed. Early zero-shot testing with standard segmentation models produced inconsistent and inaccurate delineations, confirming the need for a dedicated dataset capable of capturing the subtle morphology of spinous processes.

The CervSpineNet combines a Vision Transformer (ViT-B) encoder for global contextual modeling with a U-Net–style convolutional decoder for fine-grained boundary reconstruction (Chen J. et al., 2024; Al-Antari et al., 2025). This hybrid design maintains full spatial resolution by remaining fully convolutional, ensuring that predicted masks match input dimensions—a practical advantage for radiological workflows. Integrated residual and squeeze-and-excitation blocks selectively enhance task-relevant features and stabilize training. Collectively, these design elements allow the model to generate binary masks that closely align with ground truth and exhibit lower structural error, as confirmed across all dataset variants.

Although CervSpineNet follows the overall concept of a transformer–CNN hybrid, its architecture is specifically tailored to the demands of cervical spine X-ray segmentation. We decouple the SAM ViT-B image encoder and fine-tune it on 1,024 × 1,024 radiographs, leveraging its large-scale pretraining to capture long-range anatomical context along the full C1–C7 column. On top of these rich features, we employ a compact U-Net style decoder with residual and squeeze-and-excitation blocks, sharpening local boundaries and preserving thin, curved spinous process shapes. Training is driven by a compound loss that combines Dice overlap, Focal Tversky, Hausdorff distance–transform and SSIM terms to jointly encourage correct vertebral coverage, suppression of false positives and accurate contour geometry. In our experiments (Section 3.1, Tables 1–3), this SAM-based hybrid consistently outperforms standard U-Net, DeepLabV3+, a fine-tuned full SAM model and a text-guided SegFormer, indicating that the proposed combination of encoder, decoder and loss function confers a tangible advantage for this specific small-structure segmentation task.

Overall, the architectural ablations demonstrate that both encoder choice and decoder design substantially influence segmentation performance, albeit in complementary ways. Enhancing the decoder with residual and squeeze-and-excitation blocks consistently improves overlap, volumetric, and boundary metrics, indicating that richer decoding capacity is critical for refining fine spinal structures and suppressing noise. In contrast, replacing a purely convolutional encoder with a ViT-B transformer encoder yields the largest performance gains—particularly for boundary accuracy—highlighting the importance of long-range contextual reasoning and global shape modeling in cervical spine anatomy. Together, these results validate the design of CervSpineNet: the combination of a transformer-based encoder with an enriched decoder provides measurable advantages, while simpler components remain competitive options for lower-complexity or resource-constrained settings.

Loss ablations similarly show that region-based and structure-aware supervision outperforms Dice alone. Adding either Focal Tversky or SSIM to Dice produces consistent improvements in Dice, IoU, and SSIM, underscoring the value of class-imbalance handling and structural consistency for stable mask learning. The Dice + SSIM configuration is particularly effective for overlap and volumetric similarity, reflecting the benefit of enforcing local textural coherence in postoperative spine X-rays. These trends support our use of a compound loss: Dice and FT promote reliable region overlap and recall, SSIM preserves fine anatomical detail, and the HD95 surrogate sharpens boundaries by explicitly penalizing outlier deviations.

Across the original, CLAHE-enhanced, and augmented datasets, CervSpineNet consistently outperformed CNN-based and transformer-based baselines on Dice, IoU, SSIM, HD95, and volumetric similarity metrics. Statistical testing confirmed that these improvements were significant, highlighting the robustness and reproducibility of the approach. Notably, the model accurately segmented difficult lower cervical levels (C6–C7), which are often challenging even for experts, underscoring its ability to generalize to low-contrast or ambiguous regions.

To the best of our knowledge, this work provides the first public dataset and segmentation framework dedicated to 2D X-ray–based cervical spinous process delineation. Accurate and efficient segmentation of these structures has broad clinical value. In Anterior Cervical Discectomy and Fusion (ACDF), precise localization of spinous processes supports surgical planning and fusion evaluation (Neifert et al., 2020; Martin et al., 2025). Beyond ACDF, automatic spinous process segmentation could aid in the assessment of vertebral alignment after trauma, the monitoring of spinal deformities such as scoliosis or kyphosis, and image-guided navigation during minimally invasive procedures.

From a workflow standpoint, CervSpineNet offers substantial efficiency gains. It reduces average manual segmentation time from approximately 3–4 min per image to 5–10 s, representing a ∼96% time reduction. The lightweight architecture (∼345 MB) and CPU-compatible inference make the model suitable for clinical integration without specialized hardware. Collectively, these features position CervSpineNet as a practical, reproducible tool for advancing AI-assisted cervical spine imaging and analysis.

Despite strong performance, several limitations warrant consideration. First, training the hybrid architecture is more computationally demanding than lighter CNN baselines due to the transformer encoder and multi-stage decoding. Model performance may also be sensitive to threshold selection, class imbalance, and annotation noise, and the fixed-size image resizing may introduce subtle aspect-ratio bias. Although CervSpineNet performs well on internal data, external validation across different scanners, institutions, and acquisition protocols remains necessary to assess generalization and potential domain shift. Additionally, while we report strong segmentation metrics (Dice, IoU, HD95, etc.), we did not directly evaluate downstream clinical indices—such as spinous-process–derived angular measures or deformity parameters. Establishing links between CervSpineNet’s segmentations and clinically meaningful quantitative metrics will be essential for demonstrating translational impact.

Future optimization efforts should investigate encoder freezing, mixed-precision training, gradient checkpointing, or knowledge distillation to reduce computational burden. Boundary accuracy may be further improved through boundary-aware decoders, attention-gated skip connections, or uncertainty-based quality control mechanisms. Expanding the dataset to incorporate diverse imaging protocols, patient populations, and post-operative presentations will enhance robustness. Extending the framework to multi-view radiographs or 3D modalities (CT, MRI) may also enrich anatomical representation and support more complex clinical tasks.

Prospective clinical validation and workflow-integrated deployment studies will be critical for determining real-world utility in surgical planning, postoperative monitoring, and routine spine care. Beyond surgical settings, automated spinous-process segmentation may facilitate longitudinal deformity surveillance, trauma assessment, and AI-assisted spine navigation. These directions position CervSpineNet as a promising foundation for scalable, clinically relevant spine-imaging solutions.