Machine Learning and Deep Learning Algorithms for Skin Cancer Classification from Dermoscopic Images.

Area of Science:

• Dermatology
• Computer Science
• Artificial Intelligence

Background:

• Accurate skin tumor classification is crucial for effective treatment.
• Machine learning (ML) and deep learning (DL) offer potential for automated diagnosis.
• Existing studies show varying performance of ML and DL models.

Purpose of the Study:

• To critically assess and compare the performance of various ML and DL models for skin tumor classification.
• To evaluate the effectiveness of transfer learning using pre-trained DL models.
• To test model robustness on larger and imbalanced datasets.

Main Methods:

• Tested ML algorithms: logistic regression, linear discriminant analysis, k-nearest neighbors, decision tree, Gaussian naive Bayes.
• Employed DL models: custom Convolutional Neural Network, transfer learning with VGG16, Xception, and ResNet50.
• Evaluated models on small and large, imbalanced skin tumor datasets using metrics like accuracy and F-score.

Main Results:

• Deep learning models consistently outperformed ML models, achieving accuracies up to 0.88.
• ML models had accuracies below 0.72, improved to 0.75 with ensemble learning.
• Fine-tuned pre-trained DL models, particularly VGG16, demonstrated excellent performance on both datasets.

Conclusions:

• Deep learning models, especially VGG16 leveraging transfer learning, are highly effective for skin tumor classification.
• Transfer learning significantly enhances DL model performance for this task.
• DL models show promise for reliable and accurate automated skin tumor diagnosis.