Hugging Face's logo

Datasets:
Med-GLIP
/
Med-GLIP-5M

ArXiv:
Tags:
medical
vision-language
visual-grounding
multi-modal
pre-trained
License:
Dataset card Files
xet
Community
1
Med-GLIP-5M / README.md
Venn2024's picture
Venn2024
Update README.md
486b424 verified
preview code
|
raw
history blame contribute delete
6.96 kB
metadata 
license: cc-by-4.0
tags:
  - medical
  - vision-language
  - visual-grounding
  - multi-modal
  - pre-trained

Model Card: Med-GLIP

Model Details

  • Model Name: Med-GLIP

  • Paper Title: Med-GLIP: Advancing Medical Language-Image Pre-training with Large-scale Grounded Dataset

  • Authors: Ziye Deng, Ruihan He, Jiaxiang Liu, Yuan Wang, Zijie Meng, Songtao Jiang, Yong Xie, Zuozhu Liu

  • Affiliations: Zhejiang University

  • Version: v1

  • Date: (Presumed August 2025)

  • Model Type: Medical Language-Image Pre-training Model with Visual Grounding capabilities.

  • Relevant Links:

  • License: Creative Commons Attribution 4.0 International (CC BY 4.0)

  • Citation:

    @misc{deng2025medglip,
      title={Med-GLIP: Advancing Medical Language-Image Pre-training with Large-scale Grounded Dataset},
      author={Ziye Deng and Ruihan He and Jiaxiang Liu and Yuan Wang and Zijie Meng and Songtao Jiang and Yong Xie and Zuozhu Liu},
      year={2025},
      eprint={2508.10528},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Model Description

Med-GLIP is a medical-domain language-image pre-training model designed to enhance the understanding of fine-grained correspondences between medical images and text. In contrast to existing medical multi-modal models (e.g., MedKLIP, LLaVA-Med), Med-GLIP specifically emphasizes visual grounding, the ability to localize medical entities or findings mentioned in text to their corresponding regions in the image. The model's development is coupled with a large-scale, grounded medical language-image dataset, Med-GLIP-5M.

The model aims to overcome the limitations of existing methods in fine-grained understanding and localization, which is crucial for applications that require precise linking between report findings and image regions. By pre-training on the large-scale grounding dataset, Med-GLIP learns stronger cross-modal alignment capabilities.

Intended Use

  • Primary Intended Uses:
    • Medical Visual Question Answering (VQA)
    • Medical Report Generation (MRG)
    • Phrase Grounding: Localizing text phrases (e.g., diseases, anatomical structures) to image regions.
    • Serving as a foundational pre-trained model for various downstream medical multi-modal tasks (e.g., interactive segmentation, diagnostic assistance).
  • Primary Intended Users:
    • Medical AI researchers
    • Engineers developing medical image analysis and reporting tools
    • Researchers interested in multi-modal learning and visual grounding
  • Out-of-Scope Uses:
    • Direct use in clinical diagnostic decision-making without rigorous validation and regulatory approval.
    • Use in non-medical image-text tasks.

Training Data

  • Dataset: Med-GLIP-5M
    • A custom-built, large-scale medical language-image dataset specifically created for Med-GLIP, featuring extensive grounding annotations (correspondences between image regions and text phrases).
    • Dataset Construction: The paper details the pipeline, including Data Source Analysis, Data Collection, Data Preprocessing, Quality Control, and the generation of grounding annotations (possibly utilizing tools like SAM).
  • Composition: (Specific details depend on the full paper) Expected to include various medical imaging modalities (e.g., X-rays, CTs, MRIs) paired with corresponding radiological reports or descriptive texts, with a focus on high-quality phrase-region bounding box annotations.

Model Architecture

  • Med-GLIP is based on the architectural principles of GLIP (Grounded Language-Image Pre-training), adapted for the medical domain. Key components are expected to include:
    • Image Encoder: Likely based on a Transformer architecture (e.g., ViT or Swin Transformer) for feature extraction.
    • Text Encoder: Likely based on a BERT variant for encoding text inputs (reports and query phrases).
    • Cross-Modal Fusion Module: For deep interaction between image and text features.
    • Grounding Head: To predict bounding boxes corresponding to the input text phrases based on the fused features.
  • Training Objectives:
    • Grounding Loss: Minimizing the difference between predicted and ground-truth bounding boxes (e.g., using L1 and GIoU loss).
    • Image-Text Contrastive (ITC) Loss: Ensuring that matched image-text pairs are aligned in the feature space, facilitating global alignment. The formula is likely similar to: $L_{ITC} = -\log \frac{\exp(\text{sim}(I, T)/\tau)}{\sum \exp(\text{sim}(I, T')/\tau)} - \log \frac{\exp(\text{sim}(I, T)/\tau)}{\sum \exp(\text{sim}(I', T)/\tau)}$.

Evaluation

  • Evaluation Tasks:
    • Phrase Grounding: Evaluated on dedicated medical grounding datasets.
    • Visual Question Answering (VQA): Evaluated on standard medical VQA datasets (e.g., VQA-RAD, SLAKE, PathVQA).
    • Medical Report Generation (MRG): Evaluated on datasets like MIMIC-CXR for report quality.
  • Metrics:
    • Grounding: IoU (Intersection over Union), Recall@k.
    • VQA: Accuracy, AUC.
    • MRG: Text generation metrics such as BLEU, ROUGE, and CIDEr.
  • Results: The paper reports significant performance gains over previous state-of-the-art methods in several downstream tasks, particularly those requiring strong grounding capabilities. Figures and tables (e.g., Figure 7, Table 6) provide qualitative and quantitative comparisons.

Limitations

  • Model performance is highly dependent on the quality and coverage of the Med-GLIP-5M dataset.
  • The model's ability to generalize to rare diseases or unseen imaging modalities/styles may be limited.
  • Noise or inaccuracies introduced during the automated grounding annotation process could affect the model's precision.
  • The model's computational requirements may be high for training and inference.
  • (Refer to the full paper for a comprehensive discussion of limitations.)

Bias, Risks, and Ethical Considerations

  • Data Bias: The Med-GLIP-5M dataset may contain demographic biases (e.g., in age, gender, race representation) from its source institutions, which can be reflected in the model's performance on underrepresented groups.
  • Clinical Risk: The model is an AI research tool and must not be used for primary clinical diagnosis or patient care without explicit, strict clinical validation and regulatory approval. Misinterpretation of results could lead to patient harm.
  • Interpretability: While the grounding feature aids in interpretability, the overall decision-making process is complex, and failures should be treated with caution.
  • (Refer to the full paper for a detailed discussion of ethical and societal implications.)