Libra Model Card
Paper and Resources for More Information
For further details about Libra, including its architecture, training process, and use cases, please refer to the following resources:
- Project Website: Libra v1.0
- Article: Comprehensive paper describing Libra’s design and experiments arXiv:2411.19378
- Code Repository: Open-source implementation and pre-trained models (GitHub: X-iZhang/Libra)
Core Components:
- RAD-DINO: Vision encoder pre-trained on medical imaging datasets for robust image feature extraction.
- Meditron-7B: A large language model specialised in medical text generation, based on Llama-2.
- Temporal Alignment Connector (TAC): Custom-designed adapter for integrating temporal information between current and prior chest X-rays.
Training Strategy:
Two-stage training process:
- Temporal feature alignment.
- Fine-tuning on the radiology report generation task.
Primary Use Case:
Generates detailed findings and impressions sections for chest X-ray reports, incorporating temporal comparisons.