SpineVLM: A Markdown-Guided Structured Fine-Tuning Framework for Spine X-ray Report Generation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Natural Language Processing

Background:

• Automated medical report generation in specialized fields like spine radiography faces challenges due to data scarcity and high annotation costs.
• Existing multimodal large language models (MLLMs) often fail to detect subtle spinal abnormalities in these limited data settings.
• There is a need for efficient frameworks capable of generating accurate and structured reports for spine X-rays.

Purpose of the Study:

• To introduce SpineVLM, a data-efficient framework for structured spine X-ray report generation.
• To address the limitations of current MLLMs in detecting subtle spinal abnormalities.
• To provide a benchmark and framework for automated spine radiography report generation in data-limited scenarios.

Main Methods:

• Developed the SXRG dataset with 10,468 image-report pairs using a hierarchical AI-assisted annotation pipeline.
• Proposed Markdown-Guided Structured Learning (MGSL) to regularize learning and reformulate text synthesis into a structured completion task.
• Implemented an unsupervised Region-Focused Inference (RFI) module using DINOv2 to focus on the vertebral column for enhanced lesion detection.

Main Results:

• SpineVLM, evaluated on a 7B-parameter vision-language backbone, outperformed ten baseline multimodal models on linguistic metrics.
• Achieved a diagnostic F1-score of 0.866 in a double-blind reader study, comparable to specialist performance.
• Reduced clinical reporting time by over 41% compared to traditional methods.

Conclusions:

• SpineVLM offers a data-efficient and effective solution for automated spine X-ray report generation.
• The framework demonstrates high diagnostic accuracy and significant efficiency gains in clinical settings.
• Open-sourcing the dataset and codebase provides a valuable resource for future research and development in this area.