Melanoma Detection with Uncertainty Quantification

Created in April 15, 2025

2025   ·   Medical AI   Computer Vision   Calibration

Melanoma Detection with Uncertainty Quantification
IEEE ISBI 2025 Paper Edge Detector App GitHub Repo Live Client App
Conference Venue & Impact
IEEE ISBI 2025 International Symposium on Biomedical Imaging 📍 Houston, Texas, USA ★ Premier IEEE venue specializing in clinical vision & biomedical computation.
Paradigm Shift: Active Rejection Routing
Ambiguous Lesion 1,296 Models (10-Dataset Multi-Mix) Entropy Filter H(X) = -Σ p log(p) Certain: Accept Prediction Retain Autonomous Pipeline Uncertain: Mark "Unknown" Route Directly to Dermatologist
Realized Performance Gains & Rejection Trajectory
+4.6% Accuracy Scaled Baseline 93.2% → Optimized 97.8% -40.5% Misdiagnoses Eradicated Confident False Positives/Negatives Prevented
97.8% 95.0% 93.2% Accuracy 0.0 (No Rejection) 0.10 Threshold 0.20 (Optimal Rej) Uncalibrated: 93.2% Optimized Peak: 97.8%
Empirical Reliability Calibration (10-Bin Verification)
Perfectly Calibrated (y = x) 0.0 0.5 1.0 True Outcomes 0.0 0.5 1.0 Predicted Model Confidence Single Dataset (Overconfident) 10-Mix Multi-Backbone (Calibrated)
Curated Benchmark Breakthroughs
ISIC 2017 DenseNet201 91.9% Peak -40.5% Misdiagnoses Brier Score Optimized to 0.0478 ISIC 2018 ResNet152 95.8% Peak -66.5% False Negatives ECE Score Crushed to 0.0103 7-Point ResNet152 87.5% Peak Severe Outlier Mitigation ECE Down from 0.0396 → 0.0321 Kaggle ResNet101 97.6% Peak -81.0% False Negatives High-Risk Malignancy Routing Protected
Zero-Upload Client-Side Edge Inference
Edge Detector Instantaneous Melanoma Detector via Edge Computing 🔗 ✓ Processes inference natively inside local browser via zero file uploads.