SpineMed-450k: Multimodal Spinal Dataset

• SpineMed-450k is a multimodal dataset enabling precise vertebral-level assessment using X-ray, CT, and MRI imaging.
• It supports clinical decision-making with 456,748 instruction instances for accurate pathology grading and surgical planning.
• The dataset is curated by spine surgeons from de-identified hospital data and public sources, ensuring robust, traceable AI training.

SpineMed-450k is a large-scale, multimodal dataset constructed to advance vertebral-level reasoning and clinical decision-making in spinal disorders. Comprising approximately 456,748 instruction instances, it targets the granular assessment of pathologies and supports AI model training and benchmarking for diagnosis, pathology grading, and surgical planning based on X-ray, CT, and MRI modalities. The dataset's design, curation, and validation primarily involve practicing spine surgeons and integrate both publicly available and de-identified hospital data, ensuring a clinically grounded, traceable corpus suitable for the development of level-aware vision-LLMs (VLMs) (Zhao et al., 3 Oct 2025).

1. Clinical Foundations and Rationale

SpineMed-450k addresses the necessity of vertebral-level reasoning, fundamental for the diagnosis and management of spinal pathologies such as degenerative disc disease, fractures, deformities, infections, and tumors. Accurate integration of findings across complementary imaging modalities is essential: X-rays facilitate assessment of bone alignment and gross fractures, CT scans offer high-resolution bony detail and hardware visualization, while MRIs detect soft-tissue involvement including herniations, cord compromise, and tumors. The dataset is structured around three core clinical tasks:

• Level identification: Determining the exact vertebral segment affected (e.g., fractured, herniated).
• Pathology assessment: Grading severity and distinguishing subtypes (e.g., degree of stenosis at L4–L5, spondylolisthesis classification).
• Surgical planning and patient guidance: Recommending surgical approach, fixation levels, instrumentation strategy, and risk management.

These tasks underpin real-world clinical reasoning and guide both non-operative and operative interventions.

2. Corpus Composition and Stratification

SpineMed-450k consists of about 456,748 instruction instances, divided into 456,174 for training and 574 for testing. The dataset encapsulates four main instruction types (see Section 4) and features multimodality throughout: all major spinal imaging techniques (X-ray, CT, MRI) are represented, alongside corresponding textual sources such as guidelines, textbooks, consensus statements, and clinical notes. Case reports originate from Europe PMC, open research datasets (Spark for 2D imaging, VerSe for 3D vertebral segmentation), and approximately 1,000 de-identified hospital cases sourced from 11 centers over three years—with item-level provenance via source IDs or case IDs.

Instances are approximately uniformly distributed across 14 subconditions (cervical/lumbar degeneration, scoliosis, trauma, infection, tumor, and others), with hospital cases balanced for gender and age. The test set is exclusively composed of multiple-choice and report generation tasks, emphasizing clinical relevance and challenge.

Task Type	Train Instances	Test Instances	Total
Multiple-choice QA	248,789	487	249,276
Open-ended QA	197,413	–	197,413
Multi-turn consultations	1,138	–	1,138
Report generation	734	87	821

Table: Task stratification in SpineMed-450k (Zhao et al., 3 Oct 2025)

3. Data Generation and Validation Workflow

The data construction pipeline for SpineMed-450k employs a clinician-in-the-loop paradigm executed in five phases:

1. Dataset Collection: PDF, text, and imaging are sourced from medical textbooks, clinical guidelines, open question banks, publicly available datasets, and hospital PACS systems. Clinicians determine inclusion criteria and select diagnostically informative images and sequences.
2. Structured Information Extraction: PaddleOCR is used to convert educational PDFs into structured Markdown text, with “Picture Context Matching” algorithms aligning images with the correct narrative context.
3. De-identification and Cleaning: All personal identifiers are removed according to HIPAA standards, non-diagnostic images and tables are discarded, and a fine-grained orthopedic classifier (LLM-based) filters for spine-specific data (mapping 7 broad orthopedic classes to 14 spine sub-classes).
4. Instruction Generation (Two-Stage LLM Draft and Revision): Expert VLM models generate draft QA, dialogue, or report content based on image/text input. A revision phase applies clinician-curated criteria to ensure accuracy, level specificity, and adherence to reporting standards, with all prompt logs archived for provenance.
5. Clinician Review and Final Annotation: Board-certified spine surgeons perform quality assurance, refining wording, correcting errors, and finalizing especially test set items through sampled validation.

This curation pipeline ensures data traceability, clinical relevance, and diagnostic authenticity throughout the instruction set.

4. Instruction Types and Instance Structures

SpineMed-450k supports four instruction schema:

• Multiple-Choice QA: Clinical scenarios with associated imaging, presenting one correct answer out of four options.
  • Example:
  • Prompt: “On this sagittal T2-MRI sequence, which level shows grade III central canal stenosis?”
  • Answers: A. C4–C5, B. C5–C6, C. C6–C7, D. C7–T1 → Correct: B
• Open-Ended QA: Free-text responses to contextual image/text prompts.
  • Example: “Describe the findings on the lumbar CT axial cut at L3 and their surgical significance.”
• Multi-Turn Patient Consultations: Simulated dialogues between clinician and patient agents, informed by imaging.
  • Example:
  • Patient: “Why do I have numbness in my right thigh?”
  • Clinician: “Based on the L4–L5 disc herniation seen here …”
• Report Generation: Structured reports with six required sections (Imaging Findings, AI-Assisted Diagnosis, Treatment Recommendations, Risk & Prognosis Management, Postoperative Issue Management, Diagnostic Rationale & Disclaimer).
  • Example Output (abridged):
  • “I. Imaging Findings: Severe L4–L5 disc protrusion contacting the thecal sac…
  • II. Diagnosis: L4–L5 disc herniation causing right L5 radiculopathy…
  • III. Treatment: Patient advised microdiscectomy…”

5. Evaluation Framework and Metrics

Model performance on SpineMed-450k is measured through task-specific protocols:

• Multiple-Choice QA: Accuracy on text-only ($P_1$) and multimodal ($P_2$) items.
• Report Generation: Scored by an expert-derived rubric covering five sections, each rated 1–5 points; section scores are aggregated as follows: $$\text{Score}_{\text{total}} = \sum_{k=1}^3 w_k \cdot P_k, \quad w_k = \frac{N_k}{\sum N_i}$$ where $N_1, P_1$ = text QA, $N_2, P_2$ = image QA, $N_3, P_3$ = report tasks.
• Report Score Normalization:

$$P_3 = 20 \times \sum_{i=1}^{5} \left( \frac{1}{n_i} \sum_{j=1}^{n_i} s_{ij} \right)$$ normalized to a 0–100 scale.

• Expert Validation: LLM scoring is benchmarked against human experts using Pearson $r=0.38–0.95$ over 10 dimensions, supporting reliability as a proxy for clinician judgment.

Tasks require fine-grained localization of pathology, grading, and procedural planning at the vertebral segment level, facilitating model development attuned to true clinical reasoning.

6. Additional Dataset Characteristics and Usage

Approximately 47% of instances are spine-specific within a broader orthopedic context encompassing seven subspecialties. The 1,000 de-identified hospital cases demonstrate demographic balance and derive from 11 centers (largest contributing ~33%, smallest ~1%). Training employs curriculum learning in three stages: initializing on general medical datasets (PubMedVision, MedThoughts-8K), progressing through orthopedic non-spine content, and culminating with spine-specific QA, dialogue, and report tasks. Implementation details for model training include the Qwen2.5-VL-7B architecture, DeepSpeed zero2/zero3 optimizers, learning rates ranging from $1\times10^{-5}$ to $1\times10^{-6}$, and maximum token lengths up to 49,152.

SpineMed-450k enables the training and benchmarking of large vision-LLMs for precise, clinically relevant, level-aware reasoning in spine care. Its integration with the SpineBench evaluation framework reveals systematic weaknesses in generic VLMs when challenged by fine-grained, vertebral-level inference, and supports the development of models demonstrating significant improvements in diagnostic clarity and clinical utility as confirmed by clinician evaluation (Zhao et al., 3 Oct 2025).