Automatic Malignant and Benign Skin Cancer Classification Using a Hybrid Deep Learning Approach

Area of Science:

• Dermatology and Computational Pathology
• Artificial Intelligence in Medicine
• Digital Health Technologies

Background:

• Skin cancer incidence is rising, necessitating efficient diagnostic tools.
• Traditional skin cancer identification methods are time-consuming and costly.
• Digital technologies and machine learning offer automated classification solutions.

Purpose of the Study:

• To develop and evaluate a robust machine learning model for classifying melanoma and benign skin cancers.
• To compare the performance of different feature extraction techniques within a stacked classifier framework.
• To enhance the accuracy and reliability of automated skin cancer diagnosis.

Main Methods:

• A stacked classifier model was developed using three-fold cross-validation.
• Feature extraction was performed using deep learning models: ResNet50, Xception, and VGG16.
• The system was trained and tested on 1000 skin images, categorizing them as melanoma or benign.

Main Results:

• The proposed stacked classifier with Xception feature extraction achieved 90.9% accuracy.
• Performance was evaluated using accuracy, F1 scores, AUC, and sensitivity.
• Xception demonstrated superior performance compared to ResNet50 and VGG16 for this task.

Conclusions:

• The stacked classifier approach shows promise for accurate skin cancer classification.
• Xception is an effective feature extraction method for this dermatological application.
• Further optimization with larger datasets could lead to a highly reliable skin cancer classification system.