Medical Spine Sagittal MRI Dataset for Segmentation and Foraminal Stenosis detection

Lumbar spine disorders represent one of the most prevalent musculoskeletal conditions worldwide, particularly among the elderly population. Magnetic Resonance Imaging (MRI) is the gold-standard diagnostic tool due to its superior ability to visualize soft tissues, neural structures, and degenerative changes. However, accurate interpretation of lumbar MRI scans requires specialized clinical expertise and remains time-consuming and costly. As the demand for automated diagnostic support increases, the development of robust artificial intelligence (AI) systems critically relies on large, high-quality, and precisely annotated datasets. To address this need, we introduce a new sagittal lumbar spine MRI dataset comprising 500 patients, enriched with comprehensive anatomical annotations. The dataset includes foraminal detection labels extracted by expert neurosurgeons, provided as bounding-box annotations with clinically assigned severity grades for each lumbar level on both left and right sides. In addition, it contains pixel-level segmentation masks for the vertebrae, intervertebral discs, sacrum, and posterior elements (Posterior A and Posterior B), which were generated by an AI-based model and subsequently validated and refined by expert neurosurgeons to ensure anatomical accuracy. Both the detection and segmentation annotations are further evaluated using AI models to confirm their reliability for downstream clinical and research applications. To demonstrate the dataset's utility, we developed a complete AI-based computer-aided diagnosis (CAD) system for foraminal stenosis analysis, consisting of automated slice selection, region-of-interest localization, and severity classification. The system achieved 86% accuracy in slice selection, 90% accuracy in region of interest (ROI) localization, and 65% accuracy in severity classification, reflecting the complexity of the task and the diagnostic value of the dataset. For anatomical segmentation, the best-performing architecture, SegResNet, achieved a DSC of 97.32%, HD95 of 2.337 (mm), and a recall of 97.30%, confirming the consistency and robustness of the segmentation annotations. Overall, the proposed dataset provides a reliable, richly annotated foundation for developing and benchmarking advanced AI algorithms in lumbar spine analysis, supporting a wide range of clinical and research applications from anatomical segmentation and morphological analysis to automated foraminal stenosis assessment.