Contact Us

Turning Images into Insight

How Cogito Delivered Precision Segmentation of Vertebrae, Discs, and Canal Regions for FDA & EMA - Ready Spine AI Model
Turning Images into Insight

Challenge

The client’s spinal MRI model was trained on generic, inconsistent labels (often 2D vertebra-only masks). It failed to consistently segment vertebrae, intervertebral discs (IVDs), and the spinal canal, and missed clinically relevant findings (disc degeneration, canal stenosis, scoliosis/kyphosis/lordosis, subtle osteophytes).

Solution

We deployed a specialist-driven annotation program: radiology-trained annotators under spine-specialist supervision performed 3D structure-level segmentation, curvature/landmark tagging, and multi-stage QA (voxel-level review, IAA checks, radiologist adjudication). All work was conducted under the direct oversight of board-certified spine specialists and musculoskeletal radiologists, ensuring every dataset met clinical standards. Deliverables were provided in DICOM-seg/NIfTI with a clear ontology and data dictionary.

Outcome

Segmentation performance improved from 68% → 91% with robust generalization across MRI centers, scanners, and sequences. Manual review time dropped 40–50% via semi-automated assists. External reads showed ~95% concordance with radiologist assessments. Crucially, the high-quality, audit-ready dataset enabled the client to secure FDA & EMA clearance for their spinal AI model.

Turning Images into Insight

Background

A leading AI lab set out to build a clinically useful model for spinal MRI analysis — not just vertebra detection, but reliable segmentation of vertebrae, IVDs, and the spinal canal, with sensitivity to changes associated with degeneration, stenosis, curvature deformities, and early osteophyte formation. The goal was to support clinicians with consistent, audit-friendly masks that hold up across different MRI centers and scanner vendors.

The Challenge

Most open or legacy datasets are too narrow or inconsistent for this task. Common issues we saw:

  • Label scope gaps (vertebrae without discs or canal; 2D slices without 3D continuity).
  • Inconsistent level labeling (C/T/L levels misnumbered, L5–S1 ambiguous labeling).
  • Edge-case miss rates (mild canal narrowing, small osteophytes, endplate changes).
  • Limited generalization across scanners (field strength, sequence parameters, artifacts).

This led to unstable segmentation in production-like conditions and increased manual correction by clinicians.

Our Objectives

  • Create 3D, structure-complete labels for vertebrae, IVDs, and canal with consistent level mapping.
  • Ensure the dataset incorporated multi-sequence MRI inputs (T1, T2, STIR), recognizing that each sequence highlights different anatomical and pathological features critical to clinical reliability.
  • Improve model generalization across centers, scanners, and common sequences (e.g., T1/T2).
  • Reduce manual review time while maintaining clinical rigor and traceability.
  • Ensure the dataset was regulatory-grade, capable of supporting FDA submission.

Our solution offering

1. Data Curation & Harmonization

  • Curated a representative MRI cohort spanning multiple sites, scanners, and sequences.
  • Standardized metadata, resolved inconsistent level numbering, and excluded studies with confounding artifacts when necessary (while preserving real-world variability).

2. Structure-Level Annotation Protocol

  • Targets: Individual vertebrae (C2–L5 as available), intervertebral discs (C2–C3 through L5–S1), spinal canal (region of interest along the cervical, thoracic, and lumbar segments).
  • Granularity: 3D masks with landmark points for level confirmation and curvature mapping to support scoliosis/kyphosis/lordosis assessment.
  • Guidance: Spine-specialist-approved instructions covering partial volume, disc margins, endplate boundaries, foraminal edges, and common pitfalls (e.g., Modic changes masquerading as disc extension).
  • Formats: DICOM-SEG and NIfTI masks, plus JSON label maps aligned to a project-specific ontology.

3. Specialist-Supervised Workflow

  • Radiology-trained annotators executed primary masks; spine specialists provided decision rules for ambiguous levels and adjudicated difficult cases.
  • Semi-automated assists (smart pre-masks) accelerated routine contours; humans corrected edges and topology to clinical quality.

4. Multi-Stage Quality Assurance

  • Tier-1 QA: Voxel-level review for each mask (leakage, gaps, topology).
  • Inter-Annotator Agreement (IAA): Routine pairwise checks; discrepancies escalated.
  • Tier-2 QA: Radiologist adjudication on flagged studies and periodic blind samples.
  • Drift Control: Spot checks across centers/sequences to keep quality stable as data diversified.

5. Documentation & Hand-Off

  • Delivered a data dictionary, label ontology, and annotation guideline for reproducibility.
  • Provided ingestion scripts and example notebooks to drop the dataset into the client’s training pipeline without rework.
  • Ensured data lineage and auditability were maintained to support FDA submission requirements.

What Cogito Tech delivered

  • Segmentation performance: From 68% → 91% on key structures (vertebrae, discs, canal) across the validation cohort.
  • Efficiency: 40–50% reduction in manual correction time due to higher-quality pre-masks and consistent labels.
  • Clinical alignment: External validation showed ~95% concordance with radiologist assessments on clinically relevant regions.
  • Regulatory success: The rigor of Cogito Tech’s annotation protocols and dataset structuring enabled the client to achieve FDA clearance for their spinal MRI AI model.

Need precise annotation for your Spine AI model?

Talk to an Expert