How the Tool Works
The computer vision uses a neural network to classify images as xylazine-related or not. The developer fine-tuned a pretrained network using xylazine-related and other wound photographs collected from online sources.
The tool is an image classifier, not a large language model. This website does not outsource to a service like ChatGPT, Claude, or Gemini. Neural networks have been used in medical-imaging practice, identifying tubercolosis and retinal disease. Additional research shows how these models can assist detecting wounds via mobile application.
What the Result Means
The outputs describe model behavior stemming from the training and its underlying dataset. The outputs are not intended to establish the cause of a wound or inform diagnosis, triage, treatment, or other decisions about an individual’s care.
We do not yet know how reliably this tool will perform on real-world uploads. Results during training were encouraging but don't guarantee reliability because of how wounds and images vary. The model can be wrong in either direction. Additionally, photographs alone cannot establish the cause of a wound.
This tool is intended for researchers and learners to explore how an experimental image classifier assesses historical photos, for educational purposes. It’s a demonstration of how harm reduction can incorporate frontier technologies to enhance our abilities to assist people.
How the Tool Can Improve
Training on more photographs, both xylazine-associated and other drug-related wounds, could improve the performance of the model. Introducing a wider variety of images would help, such as different people, settings, lighting conditions, skin tones, and angles, etc.
Involving additional clinicians with experience recognizing xylazine wounds could help strengthen the underlying dataset. The model is only as good as the labels in the training data.
Other researchers and machine learning specialists may know of methods which could derive better outcomes from the training data.
Public feedback is always valuable and could inform future iterations.
Research Behind the Approach
The machine learning for this project was inspired and informed by other research:
- Sompalle et al. (2026), Development and Validation of an Image-Based Deep Learning Tool for Identification of Xylazine-Associated Wounds
- Esteva et al. (2017), Dermatologist-level classification of skin cancer with deep neural networks.
- Liu et al. (2020), A deep learning system for differential diagnosis of skin diseases.
- Fatima et al. (2026), Mind the Gap: A survey of chronic wound datasets, methods, and clinical translation.