chaoyinshe/Ultrasonic_image_desensitization

EchoVLM Ultrasound Anonymization Tool

This repository provides an ultrasound image anonymization toolkit built based on YOLOv11. It acts as the official supporting preprocessing tool for EchoVLM and EchoVLM-Thyroid-9B. The tool automatically detects and obscures sensitive text labels, patient identifiers and identifiable markers on ultrasound images to meet medical data privacy compliance requirements.

Environment Requirements

pip install ultralytics
from ultralytics import YOLO

# Load trained YOLOv11 detection model for anonymization
model = YOLO("yolo11.pt")  

# Run batch inference on thyroid ultrasound images
results = model(\["image1.jpg", "image2.jpg"\])  # Return a list of Results objects

# Process detection results and perform anonymization
for result in results:
    boxes = result.boxes          # Bounding boxes of sensitive regions
    masks = result.masks          # Segmentation masks (for segmentation models)
    keypoints = result.keypoints  # Keypoints object
    probs = result.probs          # Classification probabilities
    obb = result.obb              # Oriented bounding boxes for rotated text

    result.show()                 # Visualize detection results
    result.save(filename="result.jpg")  # Save visualized detection output
    # Implement blurring or masking logic on detected regions to generate anonymized images

Citation

If you use this tool in your research or project, please cite our EchoVLM paper:

@misc{she2026echovlmdynamicmixtureofexpertsvisionlanguage,
      title={EchoVLM: Dynamic Mixture-of-Experts Vision-Language Model for Universal Ultrasound Intelligence}, 
      author={Chaoyin She and Ruifang Lu and Lida Chen and Wei Wang and Qinghua Huang},
      year={2026},
      eprint={2509.14977},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.14977}, 
}