KUcarrot/dermassist-lmic

Offline-first AI dermatology screening for Sub-Saharan Africa | Kaggle Gemma 4 Hackathon # DermAssist LMIC > **Offline-first AI dermatology screening assistant for Sub-Saharan Africa and other low- and middle-income countries (LMICs).** In Sub-Saharan Africa, fewer than **1 dermatologist serves every 1,000,000 people**. DermAssist LMIC brings frontier dermatology AI to clinics 200 km from the nearest specialist, runs entirely offline on a single laptop, and is specifically fine-tuned for LMIC patient contexts including patients with albinism (1000x increased skin cancer risk). --- ## Demo **Video demo:** YouTube link --- ## Key Results The system was validated on two datasets to test cross-dataset robustness: | Metric | HAM10000 (in-distribution, n=35) | BCN20000 (external, n=60) | |---|---|---| | Vision Classifier accuracy | 60.0% | 28.3% | | Urgency-recommendation consistency | 100.0% | 100.0% | | Hallucination-free output | 100.0% | 98.3% | | Safety disclaimer inclusion | 100.0% | 100.0% | | **Overall safety pass rate** | **100.0%** | **98.3%** | **Key finding:** The system maintains safety guarantees under significant distribution shift, even when the upstream Vision Classifier accuracy drops by 32 percentage points. This is the result of deliberate "safety-by-design" through LMIC-specialized fine-tuning. --- ## Architecture ``` [Skin Lesion Image] | v [DullRazor Hair Removal] | v [Vision Classifier] (EfficientNet-B4) | v [Patient Context] | v [RAG Retrieval] <-- (DermNet, BAD, WHO) | v [Gemma 4 E4B + LoRA] (LMIC-specialized) | v [Urgency | Recommendation | Patient Summary | Limitations] ``` ### Components - **Vision Classifier:** EfficientNet-B4 fine-tuned on HAM10000 (10,015 dermatoscopic images, 7 classes) - **Hair Removal:** DullRazor algorithm (Lee et al., 1997) for image preprocessing - **RAG Knowledge Base:** SQLite + BAAI/bge-m3 embeddings (1024-dim, 246 chunks), indexed from DermNet, BAD guidelines, and WHO LMIC dermatology protocols - **LLM:** Gemma 4 E4B (4-bit quantized) + LoRA adapter (r=32, alpha=16) fine-tuned for …