Research map / Image generation / Editing and style transfer
B-spine: Learning B-spline Curve Representation for Robust and Interpretable Spinal Curvature Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes B-Spine, a pipeline that refines spine masks with a CycleGAN-based SegRefine network, predicts clamped B-spline centerlines from masks, and estimates Cobb angles by combining slope analysis with regression, evaluated on AASCE2019 and a new low-quality CJUH-JLU dataset.
Problem
Robust and interpretable Cobb angle estimation from low-quality, blurry spinal X-ray images, where landmark annotations are hard to obtain and centerline representations are noise-sensitive.
Approach
Represent the spine centerline as a clamped B-spline predicted from refined segmentation masks (with a CycleGAN mask-refinement step and a resample loss on curve points), then estimate Cobb angles by blending curve-slope analysis with a curve-based regression MLP.
Key results
- On AASCE2019, Ours achieves MAE of 3.73 (MT), 4.13 (PT), 4.81 (TL), outperforming Seg4Reg+ (3.88, 4.62, 4.99) and JointNet, though SMAPE (9.53) trails Seg4Reg+ (8.47).
- On CJUH-JLU, Ours achieves average-style MAE of 4.17/4.59/4.20 and SMAPE 15.55, versus CasNet 22.22 SMAPE and Seg4Reg 18.87.
- Removing SegRefine raises AASCE2019 average MAE from 4.22 to 4.64 and SMAPE from 9.53 to 10.43.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SegRefine to the initial segmentation pipeline (ablation) | B-Spine · AASCE2019 test | MAE_avg: 4.64 -> 4.22 | Helped |
| Add SegRefine to the initial segmentation pipeline (ablation) | B-Spine · CJUH-JLU test | MAE_avg: 4.39 -> 4.32 | Mixed |
| Use B-spline instead of a 6-degree polynomial (ablation) | B-Spine · AASCE2019 test | MAE_avg: 5.36 -> 4.22 | Helped |
| Use B-spline instead of a 6-degree polynomial (ablation) | B-Spine · CJUH-JLU test | MAE_avg: 5.11 -> 4.32 | Helped |
| Use B-spline instead of CubicSpline (ablation) | B-Spine · AASCE2019 test | MAE_avg: 4.65 -> 4.22 | Helped |
| Use B-spline instead of CubicSpline (ablation) | B-Spine · CJUH-JLU test | MAE_avg: 4.44 -> 4.32 | Helped |
| Use B-spline instead of CubicHermiteSpline (ablation) | B-Spine · AASCE2019 test | MAE_avg: 4.35 -> 4.22 | Helped |
| Use B-spline instead of CubicHermiteSpline (ablation) | B-Spine · CJUH-JLU test | MAE_avg: 4.49 -> 4.32 | Helped |
| Add initial-point and resample supervision to L_paras (ablation) | B-Spine CurvePred · AASCE2019 test | MAE_avg: 5.45 -> 4.22 | Helped |
| Add initial-point and resample supervision to L_paras (ablation) | B-Spine CurvePred · CJUH-JLU test | MAE_avg: 5.00 -> 4.32 | Helped |
| Add L_resample (ablation) | B-Spine CurvePred with L_init and L_paras · AASCE2019 test | MAE_avg: 4.75 -> 4.22 | Helped |
| Add L_resample (ablation) | B-Spine CurvePred with L_init and L_paras · CJUH-JLU test | MAE_avg: 4.54 -> 4.32 | Helped |
Surprising findings
- Cubic Hermite spline interpolation achieves competitive results versus B-spline on CJUH-JLU for PT angle, and the 6-degree polynomial is worse overall.
- Slope-only Cobb angles outperform regression-only angles on AASCE2019, yet the hybrid is best.
Limitations
- Multi-stage pipeline may accumulate errors from earlier modules, especially for severe low-quality inputs.
- B-spline learning is slow because the resample loss requires resampling the curve each training step.
- SMAPE metric favors larger predicted angles, making the method appear worse on that metric on AASCE2019.
Experimental setup
Datasets: AASCE2019 CJUH-JLU
Benchmarks: AASCE2019 CJUH-JLU
Metrics: MAE SMAPE IoU Dice pixel accuracy
Baselines: Seg4Reg Seg4Reg+ CasNet JointNet MDC 6-degree polynomial CubicSpline CubicHermiteSpline
Builds on: CycleGAN SSformer Seg4Reg+ CasNet Grad-CAM
Closest papers by meaning
- RouterNet: Hierarchical Point Routing Network for Robust Vertebral Landmark Localization on AP X-ray Images (AAAI 2026 · similarity 0.83)
- SpineBench: A Clinically Salient, Level-Aware Benchmark Powered by the SpineMed-450k Corpus (ICLR 2026 · similarity 0.81)
- SIMSPINE: A Biomechanics-Aware Simulation Framework for 3D Spine Motion Annotation and Benchmarking (CVPR 2026 · similarity 0.81)
- Benchmarking Encoder-Decoder Architectures for Biplanar X-ray to 3D Bone Shape Reconstruction (NeurIPS 2023 · similarity 0.81)
- Automatic Vertebra Localization and Identification in CT by Spine Rectification and Anatomically-Constrained Optimization (CVPR 2021 · similarity 0.81)
- Uncertainty modeling for fine-tuned implicit functions (ICLR 2025 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.