A comprehensive analysis of deep learning and transfer learning techniques for skin cancer classification

Area of Science:

• Dermatology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Early and accurate melanoma diagnosis is crucial but challenging due to diverse skin lesion characteristics.
• Deep learning offers promising solutions for automated skin cancer image classification.
• Existing methods require evaluation for improved diagnostic accuracy.

Purpose of the Study:

• To investigate and compare deep learning approaches for classifying dermoscopic images of skin lesions.
• To identify an effective deep learning strategy for distinguishing between benign and malignant skin cancers.
• To evaluate the performance of various pre-trained networks and machine learning classifiers.

Main Methods:

• Three deep learning strategies were evaluated: fine-tuned pre-trained networks (VGG19, ResNet18, MobileNet_V2), pre-trained networks as feature extractors with machine learning classifiers (SVM, DT, Naïve Bayes, KNN), and combined feature extraction.
• Image preprocessing included scaling, denoising, and enhancement, followed by segmentation using the active contour approach.
• The ISIC 2018 dataset, comprising 3300 skin disease images, was used for training (80%) and testing (20%).

Main Results:

• The combined approach using ResNet-18 and MobileNet_V2 as feature extractors, concatenated and fed into a Support Vector Machine (SVM) classifier, yielded the highest accuracy.
• This hybrid model achieved a maximum classification accuracy of 92.87% on the ISIC 2018 dataset.
• Different training parameters such as epoch, batch size, and learning rate were optimized for each approach.

Conclusions:

• Combining deep learning feature extraction (ResNet-18, MobileNet_V2) with SVM classification offers a highly effective method for melanoma detection.
• The developed approach demonstrates significant potential for improving the accuracy and efficiency of early skin cancer diagnosis.
• Further research can explore ensemble methods and advanced architectures for even greater diagnostic precision.