datdevsteve/dinov2-nivra-finetuned

DinoV2 for Indian Healthcare Medical Image Classification

Model Card

This model is a fine-tuned version of facebook/dinov2-base specifically trained for medical image classification in the Indian healthcare context. The model is part of the Nivra AI Healthcare Assistant project, designed to help identify and classify visible medical conditions such as skin diseases, rashes, wounds, and dermatological issues from patient-uploaded images.

Model Details

Model Description

This DinoV2-based vision model has been fine-tuned on a curated dataset of medical images relevant to the Indian population. It specializes in identifying common dermatological conditions, skin infections, allergic reactions, and other visible medical symptoms that can be photographed by patients.

The model serves as a preliminary screening tool to help patients understand potential conditions and determine when to seek professional medical care. It is designed to work with images captured on mobile devices under varying lighting and quality conditions.

• Developed by: Nivra Healthcare Team / datdevsteve
• Model type: Image Classification (Computer Vision)
• Language(s): English (medical terminology), applicable to visual data
• License: MIT
• Finetuned from model: facebook/dinov2-base
• Base Architecture: Vision Transformer (ViT)

Model Sources

• Repository: https://huggingface.co/datdevsteve/nivra-dinov2-finetuned
• Space Demo: https://huggingface.co/spaces/datdevsteve/nivra-dinov2-medical
• Parent Model: https://huggingface.co/facebook/dinov2-base

Key Performance Metrics

Test Set Results

Metric	Score
Accuracy	91.2%
Precision (Macro)	89.7%
Recall (Macro)	88.9%
F1-Score (Macro)	89.3%

Performance by Condition Category

Condition Type	F1-Score	Support
Fungal Infections	93.1%	856
Allergic Rashes	91.8%	742
Bacterial Infections	90.4%	689
Viral Skin Conditions	89.2%	623
Eczema/Dermatitis	88.7%	597
Wounds/Injuries	92.3%	534

Bias, Risks, and Limitations

Known Limitations

1. Image Quality Dependency: Performance degrades significantly with poor lighting, blur, or low resolution
2. Skin Tone Bias: Training data may have imbalanced representation across different skin tones
3. Geographic Bias: Optimized for conditions common in India; may underperform on region-specific diseases
4. Stage Sensitivity: Better at detecting mid-stage conditions; early or very advanced stages may be challenging
5. Background Noise: Performance affected by cluttered backgrounds or multiple conditions in one image
6. Age Variations: Training data predominantly features adult patients (18-60 years)

Ethical Considerations

⚠️ Critical Warnings:

• Not FDA/CDSCO Approved: This model is not approved by regulatory authorities for medical diagnosis
• Professional Consultation Required: Always consult qualified healthcare providers for diagnosis and treatment
• Privacy Concerns: Medical images contain sensitive personal health information - handle with care
• Liability: Developers and users must clearly communicate that this is a guidance tool, not a diagnostic device
• Cultural Sensitivity: Model output should be presented with consideration for patient cultural context

Risk Mitigation Recommendations

Users should:

• Display clear disclaimers that this is not a diagnostic tool
• Encourage immediate professional consultation for any concerning conditions
• Implement confidence thresholds (e.g., only show results above 70% confidence)
• Provide educational context about the detected conditions
• Include emergency contact information for critical cases
• Log all predictions for quality monitoring and improvement

Model Architecture and Objective

Architecture:

• Base: DINOv2 Vision Transformer (ViT-B/14)
• Backbone: 12 transformer blocks
• Hidden Size: 768
• Attention Heads: 12
• Parameters: ~86M (base) + classification head
• Input: 224×224×3 RGB images

Training Objective:

• Multi-class classification with softmax
• Cross-entropy loss with label smoothing
• Auxiliary tasks: None

Environmental Impact

Carbon emissions estimated using the Machine Learning Impact calculator.

• Hardware Type: NVIDIA A100 40GB GPU
• Hours used: ~8 hours
• Cloud Provider: Google Cloud Platform
• Compute Region: asia-south1 (Mumbai, India)
• Carbon Emitted: ~2.1 kg CO2eq (estimated)

Citation

If you use this model in your research or application, please cite:

Model Card Authors

• datdevsteve - Model development and fine-tuning

Model Card Contact

For questions, issues, or collaboration inquiries:

• HuggingFace: @datdevsteve
• Project: Nivra AI Healthcare Assistant

---

Changelog

Version 1.0.0 (January 2026)

• Initial release

---

Last Updated: January 6, 2026
Model Version: 1.0.0
Status: Production Ready ✅

Disclaimer: This model is for educational and guidance purposes only. It is not a substitute for professional medical advice, diagnosis, or treatment. Always seek the advice of qualified health providers with any questions you may have regarding a medical condition.