Inspiration
Access to dermatology care remains a major challenge for many communities. Long wait times, limited specialist availability, and geographic barriers can prevent patients from receiving timely guidance for skin conditions.
We wanted to build a tool that helps bridge this gap by providing an accessible way for patients and healthcare providers to evaluate skin concerns earlier. Our goal was not to replace dermatologists, but to create an AI-assisted system that can help prioritize cases and provide preliminary insights when specialist access is limited.
What it does
PaliDerm is an AI-powered teledermatology platform that allows users to submit a skin image along with relevant symptoms and patient information. The system analyzes this multimodal input and generates ranked dermatological condition categories to assist with preliminary triage.
Rather than forcing a single prediction, PaliDerm provides multiple possible outcomes, helping communicate uncertainty and supporting better decision-making.
The platform is designed as a decision-support tool to help connect patients with appropriate care faster.
How we built it
PaliDerm combines a full-stack healthcare workflow with a multimodal machine learning pipeline.
Application
- Built a web-based teledermatology interface for submitting patient cases
- Created a backend pipeline for securely handling patient information and uploaded images
- Integrated AI inference directly into the case submission workflow
Machine Learning
The AI system was trained using the SCIN dermatology dataset.
We developed a multimodal architecture combining:
- Computer vision model: Extracts visual features from skin images
- Metadata model: Processes patient-provided information and symptoms
- Feature fusion layer: Combines image and metadata representations to generate predictions
We experimented with both transfer learning approaches and custom convolutional neural network architectures to improve performance while maintaining lightweight deployment capabilities.
Challenges
Medical datasets present unique challenges compared to traditional image classification problems.
One of the largest challenges was handling severe class imbalance. Common dermatological conditions may have thousands of examples, while rare conditions may have very limited data.
To address this, we grouped related conditions into broader clinically meaningful categories and focused on top-k prediction performance rather than relying only on single-label accuracy.
We also prioritized responsible AI design by presenting results as assistance for triage rather than definitive medical diagnoses.
Accomplishments
We successfully built an end-to-end AI-assisted teledermatology prototype connecting:
- Patient case submission
- Image processing
- Symptom analysis
- Multimodal AI inference
- Ranked condition recommendations
The system demonstrates how lightweight AI tools can help improve access to dermatological guidance, especially in environments where specialist availability is limited.
What we learned
Building healthcare AI requires more than maximizing model accuracy.
We learned the importance of:
- Managing real-world dataset imbalance
- Designing models around uncertainty
- Evaluating multiple performance metrics
- Creating responsible interfaces for sensitive applications
A useful healthcare AI system must balance technical performance, usability, and trust.
What's next
Future improvements for PaliDerm include:
- Expanding training data for rare conditions
- Improving model calibration and uncertainty estimation
- Incorporating clinician feedback loops
- Adding referral prioritization based on severity indicators
- Expanding deployment options for low-resource healthcare environments
Built With
- Python
- TensorFlow / Keras
- Flask
- SQLite
- SCIN Dermatology Dataset
- Computer Vision
- Multimodal Machine Learning
Log in or sign up for Devpost to join the conversation.