Precision AI for Early Dermatological Screening

SkinScan.io is a high-performance diagnostic support tool built for the BACSA Hacks 1-Day Challenge at the University of Toronto. It addresses the "Open Challenge: Disease Detection" by transforming mobile imagery into actionable clinical data.

By combining a specialized Convolutional Neural Network (CNN) for lesion classification with an Agentic AI reasoning layer, SkinScan.io identifies potential risks and provides structured justifications for its findings, bridging the gap between raw data and patient understanding.

• Neural Lesion Detection: Real-time analysis of skin anomalies using a CNN trained on 2,000+ clinical images.
• Unified Risk Index (URI): A custom mathematical model that weights prediction probability against model confidence to provide a transparent risk level.
• Longitudinal "Mole Tracking": A secure history dashboard that allows users to monitor the evolution of specific spots over time—crucial for identifying malignant growth.
• Agentic Clinical Reasoning: An LLM-powered engine that translates technical metadata (asymmetry, border irregularity) into human-readable reports.

A core requirement for this challenge was to model uncertainty and justify assumptions. SkinScan.io does this through our proprietary Risk Index formula:

$$URI = P(\text{condition} \mid \text{visuals}) \times C(\text{detection confidence})$$

Where:

• $P$ is the softmax probability from the final layer of the CNN.
• $C$ is the bounding-box confidence score reflecting image quality and lighting.

If the $URI$ falls below a defined threshold, the system triggers a "Low Confidence" state, advising the user to retake the photo under better lighting rather than providing a high-variance guess.

• Node.js v18+
• Python 3.9+
• Roboflow API Key

git clone https://github.com/yourusername/skinscan-io.git
cd skinscan-io

cd backend
pip install -r requirements.txt
# Create a .env file with your ROBOFLOW_API_KEY and SUPABASE_URL
python main.py

cd frontend
npm install
npm run dev

• Clinical Responsibility: We learned that a "High Confidence" prediction on a "Low Quality" image is a failure state. We architected a validation layer to ensure data integrity.
• Data Normalization: Handling variations in skin tones and lighting was the biggest hurdle. We implemented pre-processing steps to normalize contrast before inference.
• Asynchronous Pipelines: Orchestrating vision APIs and reasoning LLMs concurrently required robust state management to keep the UI snappy.

1. On-Device Inference: Porting the model to TensorFlow Lite for 100% offline, private scanning.
2. Fitzpatrick Skin Type Integration: Enhancing the model to better account for diverse skin melanin levels.
3. Telehealth Integration: One-click sharing of "Mole Tracking" history with board-certified dermatologists.