Tier 2How Solo Dermatologists Can Use AI · A Three-Tier Practical Framework

Local AI for Dermatology

Open-Source Model Deployment, Fine-Tuning, and Verified Air-Gap on Consumer Hardware
Amber Abir MD¹ · Prof. Baruch Kaplan MD⁵ · Yehonatan Kaplan MD²,³,⁴
¹ Sheba Medical Center, Tel-Hashomer · ² Maccabi Healthcare Services · ³ Assuta Medical Center · ⁴ Herzliya Medical Center · ⁵ Adelson School of Medicine, Ariel University · dr@kaplanderm.com
Local AI · The Tier 2 Approach

Cloud AI offers contractual privacy; local deployment on consumer hardware delivers architectural privacy. Open-source LLMs and AI-optimised laptops make a fully local clinical AI workstation feasible for solo dermatology. DERMind, the reference implementation, integrates four on-device capabilities: multilingual voice transcription, dermoscopy analysis, structured clinical assessment, and verified air-gap status.

Figure 1 · DERMind Local Workstation · Multi-Modal On-Device Dashboard

DERMind v1.0

🛡 Air-Gap Active ✕ No Internet
🎙 Voice Transcription Whisper v3 · NPU
Session 0:42 ▲ 98.4% accuracy
נודול עם טלאנגיאקטזיות על גב האף, קוטר כ-8 מ״מ, קיים מזה כשנה עם גדילה איטית
Nodule with telangiectasias on the nasal dorsum, ~8 mm diameter, present for ~1 year with slow growth.
🇮🇱 Hebrew 🇬🇧 English 🇸🇦 Arabic 🇷🇺 Russian
📚 Medical-term correction טלאנגיאקטזיה ✓ דרמוסקופיה ✓ + 4 more
🔬 Dermoscopy Analysis MedGemma + PanDerm
BCC (Nodular) High Risk
89.2% confidence · ensemble vote 4/5
📏 8.0 × 6.4 mm 📍 Nasal dorsum · H-zone
Differential
Melanoma6.2%
SCC3.8%
Dysplastic1.5%
Arborizing vessels Blue-gray ovoid nests Ulceration Leaf-like areas Spoke-wheel Shiny white structures
📋 Clinical Assessment MedGemma 27B · LoRA-tuned
Nodular BCC on the nasal dorsum. H-zone anatomical location with poorly defined lateral margins; size and depth indeterminate on clinical exam alone.
⚠ High-risk: H-zone 📊 ICD-10: C44.319 📅 Onset: ~12 mo
Recommendations · Personalised Safety Rules
Mohs micrographic surgery within 6 weeks
Pre-op CBC and coagulation if on anticoagulant
Histopathology confirmation pre-procedure
Follow-up dermoscopy at 3 and 12 months
Sun-protection counselling and full skin exam q6mo
🔒 Security Status
✓
ISOLATED
All models running on-device · Uptime 247 days
WiFiDisabled
EthernetDisabled
BluetoothDisabled
CellularN/A
MedGemma 27B Whisper v3 PanDerm Consultation
RTX 5090 24GB VRAM 64GB RAM
HIPAA ISO 27001 Israeli Privacy Law
⊘ Zero Data Egress · Wireshark verified

Background & Methods

Cloud AI introduces inherent dependencies: connectivity, recurring costs, vendor lock-in, contractual privacy. For dermatologists handling dermoscopic images and surgical documentation, fully local AI provides the strongest architectural privacy guarantee: zero PHI egress.

Narrative review through May 2026: open-source LLMs, consumer GPU hardware, fine-tuning, multilingual transcription, and air-gap verification.

DERMind · Four On-Device Capabilities

  • Clinical note generation: structured documentation via MedGemma 27B
  • Multilingual transcription: Hebrew, Arabic, English, Russian via Whisper large-v3 on NPU
  • Dermoscopy decision support: MedGemma 1.5 4B multimodal; LoRA fine-tuning on ISIC
  • Verified air-gap: firmware disablement plus Wireshark capture confirming zero egress

Open-Source Models · May 2026

ModelWhy it matters
MedGemma 1.5Trained on derm, histopath, CXR (Jan 2026)
Gemma 4 31B / 26B-A4BHebrew and Arabic; 256K context
Llama 4 Scout 8B10M-token context window
DeepSeek R1 32BReasoning-tuned; MIT licence
Mistral Small 24B97.8% of GPT-4.1 on PHI extraction

Conclusions

A local AI infrastructure for solo dermatology is now achievable using consumer hardware, open-source models, and freely available frameworks, without cloud dependency, recurring costs, or programming expertise beyond basic command-line operation.

This review provides a practical roadmap to the strongest privacy architecture currently available: zero PHI egress, verified by air-gap operation.

Added value: dermoscopy decision support and Hebrew transcription that never leave the device, feasible for solo practice for the first time.