Melanoma Detection with Uncertainty Quantification

TL;DR

This work combines publicly available datasets to enhance data diversity, allowing numerous experiments to train and evaluate various classifiers, and calibrates them to minimize misdiagnoses by incorporating uncertainty quantification.

Abstract

Early detection of melanoma is crucial for improving survival rates. Current detection tools often utilize data-driven machine learning methods but often overlook the full integration of multiple datasets. We combine publicly available datasets to enhance data diversity, allowing numerous experiments to train and evaluate various classifiers. We then calibrate them to minimize misdiagnoses by incorporating uncertainty quantification. Our experiments on benchmark datasets show accuracies of up to 93.2% before and 97.8% after applying uncertainty-based rejection, leading to a reduction in misdiagnoses by over 40.5%. Our code and data are publicly available, and a web-based interface for quick melanoma detection of user-supplied images is also provided.

Figures (5)

• Figure 1: Our framework includes aggregating diverse datasets, optimizing predictions through 1,296 experiments with 54 combinations of combined datasets and 24 CNNs, filtering uncertain predictions with entropy scores, referring ambiguous cases for human evaluation, and enhancing classification performance.
• Figure 2: We developed a web application that allows users to estimate a melanoma risk score of user-supplied images using web-based inference without upload (edge computing).
• Figure 3: Accuracy improves with varying rejection thresholds on validation sets, commonly in the range of 0-0.2.
• Figure 4: Calibration curves of ResNet50 as a demonstration. Left: Single dataset (ISIC'20). Right: Combined datasets (ISIC'16, ISIC'17, ISIC'18, MEDNODE, Kaggle).
• Figure 5: A plot comparing false diagnoses before and after applying uncertainty-based rejection across benchmarks.