PelliScope: Intelligent Skin Health with AI
Attention-Based Multi-Instance Learning for Real-World Dermatology
Background: A Critical Gap in Dermatology AI
Standard AI
Single image inputs
GAP
close-up
context
angled
Clinical Reality
Multi-photo cases
Objective: A Compute-Efficient, Case-Level Pipeline
Aggregate 1-3 photos per case via Multiple Instance Learning (MIL).
Leverage frozen dermatology foundation embeddings for efficiency.
Deliver tunable operating points for 10 common conditions.
Diagnostic Classes:
Eczema, Allergic & Irritant Contact Dermatitis,
Insect Bite, Urticaria, Psoriasis, Folliculitis, Tinea, Herpes Zoster,
& Drug Rash.
Methods: Comprehensive Evaluation of Multiple Instance Learning Model
Input: Proprietary Dataset
Client's proprietary dataset
10 diagnostic skin lesion classes
Step 1: Feature Extraction
Frozen Derm Foundation Encoder
(vs. ResNet, ViT, CLIP, etc.)
Step 2: Case-Level Learning
Gated-Attention MIL Head
(Trained with AdamW, cb-BCE)
Output
Case-Level Diagnosis
(Micro/Macro AUC)
Results: Outperforms Open & Closed Models
Micro ROC-AUC for 10 Diagnostic Classes
Grok-4
0.702
GPT-5
0.81
Gemini 2.5
0.82
CLIP + RF
0.802
Derm F. + G. Boost
0.841
PelliScope
0.863
+0.022
Conclusion: A Practical Path to Deployable AI
High Overall Accuracy: Achieves mean accuracy of 0.802 on the test set.
Clinically Meaningful: Tunable thresholds for specific needs.
e.g., Urticaria: Sensitivity 0.818, Specificity 0.826
Efficient & Deployable: Lightweight model on frozen features.
~6–7s inference on a consumer CPU
The solution is acknowledged by Emirates Health Services (UAE)
to be implemented in their workflow.
(Acknowledgement Letter)