Classification of Skin Lesion through Active Learning Strategies

Area of Science:

• Dermatology and Computer Vision
• Medical Image Analysis
• Machine Learning for Healthcare

Background:

• Melanoma and non-melanoma skin cancers are frequent in Brazil, with melanoma showing higher lethality.
• Deep learning for skin cancer diagnosis requires extensive annotated data and computational resources.
• Active learning strategies can optimize classifier training by selecting informative data subsets.

Purpose of the Study:

• To explore active learning approaches for efficient skin lesion classification.
• To identify optimal data selection criteria for training effective skin cancer diagnostic models.
• To reduce the computational cost and annotation burden in skin cancer diagnosis.

Main Methods:

• Extensive experimental evaluation across three datasets using various learning strategies.
• Comparison of U-net CNN and Fully Convolutional Networks (FCN) for image segmentation, including manual expert review.
• Analysis of handcrafted and deep features, classifiers (e.g., Random Forest), and active learning criteria (e.g., Margin Sampling, Entropy).

Main Results:

• Fully Convolutional Networks (FCN) with manual correction, Border-Interior Classification (BIC) extractor, and Random Forest (RF) classifier demonstrated superior performance.
• The Margin Sampling active learning strategy achieved approximately 93% accuracy using only 35% of the training data.
• Active learning significantly reduced the required training set size compared to traditional methods.

Conclusions:

• Active learning strategies enable high accuracy in skin lesion diagnosis with fewer labeled samples and faster training.
• The findings suggest active learning can substantially aid in skin lesion diagnosis, lowering annotation costs for specialists.
• Optimized data selection through active learning enhances the efficiency and effectiveness of skin cancer diagnostic tools.