Step 1 of 4
The Problem
Industry: Digital Health -- AI-Assisted Dermatology
Company Profile: DermaScan AI
DermaScan AI is a Series B digital health startup (founded 2022, 85 employees) that provides AI-powered skin lesion analysis to dermatology clinics and primary care practices across the United States and Europe. Their core product is a cloud-based diagnostic support tool that analyzes dermoscopic images (high-magnification photographs of skin lesions) and provides clinicians with risk assessments for melanoma and other skin cancers.
The Business Challenge
Skin cancer is the most common cancer worldwide, with over 5 million cases diagnosed annually in the United States alone. Early detection of melanoma -- the deadliest form -- improves 5-year survival from 30% (late-stage) to 99% (early-stage). However, dermatologist availability is severely limited: there are approximately 12,000 practicing dermatologists for 330 million Americans, creating wait times of 4-8 weeks in many regions.
DermaScan's current system uses a ResNet-50 encoder (pretrained on ImageNet) fine-tuned on 25,000 labeled dermoscopic images. The system classifies images into 7 diagnostic categories (melanoma, basal cell carcinoma, benign keratosis, dermatofibroma, melanocytic nevus, vascular lesion, actinic keratosis).
The problem: The current system achieves 82% overall accuracy, but only 74% sensitivity on melanoma -- meaning 26% of melanoma cases are missed. The clinical requirement is >90% melanoma sensitivity (to serve as a reliable screening tool) with a false positive rate below 15% (to avoid unnecessary biopsies).
Stakes
- Clinical impact: Missing a melanoma diagnosis can be fatal. Every percentage point of sensitivity improvement translates to hundreds of saved lives annually across their 200+ clinic partners.
- Revenue impact: DermaScan operates on a per-scan pricing model ($4.50/scan). With 500,000 scans/month, achieving the clinical accuracy threshold unlocks regulatory approval (FDA 510(k) pathway) and a projected 3x expansion in clinic adoption.
- Competitive pressure: Three competitors are developing similar tools. First to achieve FDA clearance captures the market.
Constraints
- Labeled data: 25,000 dermoscopic images with histopathology-confirmed labels. Acquiring new labeled data costs $85 per image (requires biopsy confirmation).
- Inference latency: Clinicians expect results within 3 seconds per image (including upload, processing, response).
- Compute budget: Inference must run on a single NVIDIA T4 GPU (clinic-deployed edge devices).
- Regulatory: The model must provide interpretability (attention maps or saliency) to satisfy FDA requirements for clinical decision support.