mehultyagi/cutisight

Cutisight: An Experimental Dermatology-Specific Vision-Language Model

Model Overview

• Cutisight is an experimental vision-language model specifically designed for dermatological applications.
• Built on a lightweight, domain-specific architecture inspired by LLaMA-based language models and CLIP.
• It is optimized for edge devices like smartphones and offline use.
• The model aims to provide accurate diagnostic assistance and educational support in dermatology while ensuring data privacy and security.
• Cutisight represents a significant step forward in domain-specific AI applications by combining efficiency, accessibility, and privacy-focused design, offering a practical tool for advancing dermatological care and education.

Model Name: Cutisight
Version: 1.0 (Experimental)
Developer: Mehul Tyagi
License: Proprietary License (Access by Permission Only)
Access: Gated Model on Hugging Face (Request via [email protected])

Capabilities

• Skin Condition Classification: Identifies and categorizes various dermatological conditions from clinical images.
• Medical Report Generation: Produces patient-friendly diagnostic summaries based on visual input.
• Interactive Recommendations: Offers treatment suggestions tailored to the input data.
• Educational Insights: Assists in dermatology education by providing explainable AI-driven insights.

Technical Specifications

• Architecture: Hybrid vision-language model
• Inputs: Clinical skin images, textual medical queries, or combined image-text data.
• Outputs: Diagnostic classifications, detailed condition explanations, and further recommendations.
• Image Processing:
  • 512x512 image patches
  • 64 visual tokens per patch
• Memory Optimization:
  • Radical image compression
  • Patch-based processing
• Special Tokens: Custom subimage division tokens

Training Data

The model is trained on a modified version of the Fitzpatrick17k and a private repository dataset, which includes diverse skin tone representations and dermatological conditions. This ensures robust performance across varied patient demographics.

Key Advantages

1. Small Size and Efficiency: Cutisight is compact and resource-efficient, making it ideal for deployment on edge devices with limited computational power.
2. Offline Functionality: The model operates entirely offline, ensuring that no sensitive patient data is sent to servers, thereby prioritizing privacy and data security.
3. Accessibility: Freely available upon eligible requests via email at [email protected], enabling controlled access for clinical and educational purposes.
4. Optimized for Edge Devices: Designed to work seamlessly on mobile devices or other low-resource hardware without compromising performance.

Intended Use Cases

• Clinical decision support for dermatologists.
• Patient education and engagement in understanding skin conditions.
• Telemedicine applications in dermatology.
• Research in dermatological AI tools and applications.

Limitations

• The model is still under training; evaluation metrics are awaited.
• Performance on rare or highly atypical conditions may be limited due to dataset constraints.
• Users should consult licensed medical professionals for final diagnoses and treatment decisions.

Ethical Considerations

Cutisight emphasizes ethical AI use by ensuring patient privacy through offline functionality and restricted access via a proprietary license. Its deployment is designed to comply with data protection standards, making it a secure tool for sensitive medical applications. However, users are advised not to rely solely on the model for critical medical decisions without professional oversight.

Access Instructions

The model is hosted as a gated repository on Hugging Face. To request access:

1. Log in to your Hugging Face account.
2. Visit the Cutisight repository page.
3. Submit an access request with your details and intended use case.
4. Await manual approval from the developer.

Note: Cutisight is under active development. Performance may vary across different skin types and conditions. Always consult a medical professional for clinical decisions.

License

1. Cutisight Proprietary License Copyright (c) 2024-2025: Mehul Tyagi

• Access is restricted and requires explicit approval
• Contact: [email protected] for access requests
• License URL: https://huggingface.co/mehultyagi/cutisight/blob/main/LICENSE.md

2. Key restrictions:

• Non-transferable, non-exclusive, and revocable license
• No redistribution, replication, or modification allowed
• Commercial use requires separate licensing agreement
• Proper attribution required for academic/research use

Permission for use of base model

Copyright (c) 2012-2021: -Gabriel Ilharco -Mitchell Wortsman -Nicholas Carlini -Rohan Taori -Achal Dave -Vaishaal Shankar -John Miller -Hongseok Namkoong -Hannaneh Hajishirzi -Ali Farhadi -Ludwig Schmidt

Permission is granted free of charge to: -Use, copy, modify, publish, distribute, sublicense, and/or sell the Software -Include this copyright notice and permission notice in all copies

Note: These are two separate licenses - Cutisight has a restrictive proprietary license while Open_CLIP uses a more permissive open-source license.