LSMA-PQR: Lumbar Spine Multi-view Annotations with Pfirrmann grading, Quantitative measurements and structured Radiological reports
Description
A clinically curated multi‑view lumbar MRI dataset for segmentation, grading, and quantitative analysis. Contains T1‑ and T2‑weighted axial and sagittal scans for 515 patients (4,120 images total: 3,090 axial, 1,030 sagittal) at 384×384 resolution. Each image is annotated for seven anatomical structures across complementary imaging planes: axial view captures intervertebral discs (IVD), posterior elements (PE), and spinal canal (SC) at three clinically relevant levels (L3-L4, L4-L5, L5-S1), while sagittal view depicts vertebral bodies, intervertebral discs, sacrum, and longitudinal spinal canal in mid-sagittal cross-section. Quantitative degeneration assessment includes: - 1,545 radiologist-validated Pfirrmann grades (515 patients × 3 disc levels) representing standardized MRI-based disc degeneration severity (Grade 1-5 scale) - 1,545 intervertebral disc height measurements with pixel-coordinate provenance (mean: L3-L4 = 10.32 ± 1.87 mm; L4-L5 = 10.15 ± 2.14 mm; L5-S1 = 8.98 ± 2.45 mm) - Structured clinical metadata extracted from 515 radiologist reports following systematic error correction (340 transcription/semantic errors corrected across 59% of reports) Features - The dataset enables supervised training for anatomical structure delineation with 100% annotation completeness. Researchers should note that class IDs are view-specific and non-overlapping (axial: 0-2; sagittal: 3-6), requiring conditional class masking in multi-view fusion architectures to prevent cross-view misclassification. - Pfirrmann grades represent ordinal categorical data (1 < 2 < 3 < 4 < 5) where prediction errors should be penalized proportionally to grade distance (predicting Grade 2 when truth is Grade 5 is clinically worse than predicting Grade 4). - The dataset supports structure-symptom analyses by linking pixel-level anatomy to narrative pathology mentions (disc bulge: 53.7%; compression: 78.3%; herniation: 13.9%) and severity categorizations (mild: 34.3%; moderate: 22.8%; severe: 19.1%). - It also supports co-registration of segmentation masks, degeneration grades, disc heights, and clinical descriptors. Standardized Filenames - Axial: {Modality}{PatientID}{DiscLevel}.png (e.g., T1_0042_D4.png) - Sagittal: {Modality}{PatientID}_S{SliceNumber}.png (e.g., T2_0001_S8.png) Dataset Contents - Images Folder : Containing 4120 Images (3090 x Axial and 1030 x Sagittal Images in .png format) - Labels in YOLO format. (4,120 YOLO .txt annotations) - Masks (4,120 semantic segmentation masks) - Overlay Visuals (4,120 quality control overlays (PNG)) - Pfirrmann Grading (.csv file) - IVD Heights (.csv file) - Radiological Notes (.csv file) Our Dataset is free to use with a request to cite our work.
Files
Steps to reproduce
Step - 1 : Extraction of Multiple Axial Slices (D3, D4, D5) and Mid-Sagittal View in Sagittal Plane for both T1-weighted and T2-weighted Images. Total Image Count is 4120. Resolution: 384×384 pixels, 0.6875 mm/pixel spacing, Per patient: 6 axial slices (T1/T2 × 3 levels) + 2 sagittal slices (T1/T2 × 1 mid-plane) [Images extracted from Original Dataset by Sudirman et al. https://data.mendeley.com/datasets/k57fr854j2/2] Step - 2 : Using MATLAB Image Labeler, perform Annotations. 3 - Classes for Axial and 4 for Sagittal View. Annotation is performed and validated by Radiologists. Manual annotation effort required for a. Axial: 3.3 min/image (1.1 min/object × 3 objects) b. Sagittal: 18.5 min/image (1.2 min/object × 15.4 objects) c. Total effort: Approximately 550 person-hours + 62 hours QC Step - 3 : Comparison of Qualitative Pfirrmann Grading with Dataset specific quantitative Pfirrmann Grading and storing the values in .csv file Patient ID wise. Coverage: 1,545 disc assessments (515 patients × 3 levels). Step - 4 : Computation of IVD Heights and storing in .csv file (In total 1545 Gradings) Step - 5 : Correction of Radiological Notes to structured error free notes. (515 x IDs)