Investigation of Binary and Multiclass Classification Performance of Skin Cancer Images Using Transfer Learning...

Area of Science:

• Dermatology
• Computer Science
• Artificial Intelligence

Background:

• Skin cancer classification is a critical diagnostic challenge.
• Transfer learning offers a promising approach for improving diagnostic accuracy.
• Optimizing pre-processing and hyperparameter tuning is essential for deep learning model performance.

Purpose of the Study:

• To evaluate the effectiveness of transfer learning models for skin cancer classification.
• To compare different learning architectures and their performance.
• To investigate the impact of pre-processing techniques and hyperparameter tuning on model accuracy.

Main Methods:

• Utilized binary and multiclass classification tasks with International Skin Imaging Collaboration (ISIC) datasets.
• Applied pre-processing techniques including DullRazor, Histogram Equalization, and Gamma Correction.
• Employed data augmentation, early stopping, and learning rate reduction for optimization.

Main Results:

• The Visual Geometry Group 16 (VGG16) model achieved the highest accuracy (0.9017 for binary, 0.9292 for multiclass) after hyperparameter tuning.
• ResNet50 and DenseNet121 also showed strong performance, demonstrating the value of optimization.
• Pre-processing and hyperparameter tuning critically enhanced model accuracy.

Conclusions:

• Transfer learning models, when combined with pre-processing and tuning, are highly effective for skin cancer classification.
• The VGG16 model shows significant potential for integration into dermoscopy systems.
• Further research should focus on diverse datasets and refined pre-processing methods.