1,000 lumbar spine MRI studies annotated for disc herniation, foraminal stenosis, and Modic changes

Overview

We are assembling a comprehensive lumbar spine MRI dataset to train radiomics and deep-learning models for automated grading of degenerative disc disease, disc herniation morphology classification, and neural foraminal stenosis severity assessment. Required cases include adult patients aged 25–80 years presenting with low back pain, radiculopathy, neurogenic claudication, or pre-operative surgical evaluation, scanned on 1.5T or 3T systems using a posterior phased-array spine coil. Field strength of 3T is preferred for its superior soft tissue contrast and improved nerve root visualisation within the neural foramen, but well-performed 1.5T studies with adequate SNR are fully acceptable. Mandatory sequences per study: sagittal T1-weighted with TR 400–700 ms, TE 10–15 ms, slice thickness ≤4 mm, FOV 250–300 mm, and matrix ≥256×256; sagittal T2-weighted with TR 3000–5000 ms, TE 80–120 ms and matching slice geometry to T1 for direct co-registration; and axial T2-weighted images acquired at L3-4, L4-5, and L5-S1 disc levels with slice thickness ≤4 mm and in-plane resolution ≤0.6 × 0.6 mm. Optional sequences valued for enhanced scientific utility include sagittal STIR for bone marrow oedema and inflammatory endplate disease; sagittal T2 fat-suppressed for epidural and ligamentous pathology; and post-gadolinium T1-weighted sagittal and axial sequences for post-operative epidural fibrosis assessment in revision surgery cases. DICOM is the required primary delivery format with all sequence parameters preserved in the DICOM header; NIfTI with JSON acquisition sidecars containing TR, TE, flip angle, slice thickness, and field strength is accepted as an equivalent alternative for institutions using BIDS-structured archives. Structured annotations must be provided per intervertebral disc level from L1-2 through L5-S1 covering the following elements: Pfirrmann degeneration grade I through V on sagittal T2 based on nucleus pulposus signal intensity and disc height; disc herniation morphology classified as none, annular bulge, focal protrusion, broad-based protrusion, extrusion with or without cranial or caudal migration, or sequestration using the 2014 NASS nomenclature; herniation zone classified as central, right or left paracentral, right or left foraminal, or far lateral; Modic endplate change type 0, I, II, or III bilaterally at each level; and neural foraminal stenosis grade 0 through 3 bilaterally at each level based on obliteration of the perineural fat signal. All per-level structured annotations must be delivered as JSON objects keyed by disc level and patient study identifier. Segmentation masks of herniated disc material on the axial T2 slices at the most affected level are optional and compensated at a 20% premium per annotated level. Bounding boxes around the primary herniation site on the most diagnostically informative axial slice are required for all extrusion and sequestration cases. Coronal reformations for scoliosis measurement are optional. De-identification must encompass removal of all 18 HIPAA-defined identifiers including patient name, date of birth, medical record number, accession number, device serial number, and geographic subdivisions smaller than state; scan dates must be shifted by a uniform per-patient random integer offset in the range of 1–365 days to prevent date-based triangulation. The intended clinical application is a decision support tool for spine surgeons, pain management physicians, and physiatrists that automatically generates structured disc-level reports from raw lumbar MRI input, reducing inter-radiologist reporting variability and turnaround time in high-volume practices. Post-operative cases with metallic fusion hardware may be included only if the artefact does not obscure the annotated disc levels and must be flagged with implant type in metadata.

Progress

Data Specifications

Use Cases

• Training and validating Medical imaging AI/ML models
• Benchmarking Medical imaging detection and segmentation algorithms
• Building de-identified Medical imaging research datasets for academic studies
• Augmenting existing Medical imaging datasets to reduce class imbalance