Hybrid of Deep Feature Extraction and Machine Learning Ensembles for Imbalanced Skin Cancer Datasets.

Area of Science:

• Dermatology
• Medical Imaging
• Computational Biology

Background:

• Skin cancer is a prevalent and dangerous disease requiring precise early diagnosis.
• Existing classification methods struggle with imbalanced datasets, hindering accurate diagnosis.

Purpose of the Study:

• To develop an advanced skin cancer classification model using combined machine learning (ML) and deep learning (DL) techniques.
• To enhance classification performance on imbalanced skin cancer datasets.

Main Methods:

• Feature extraction from deep learning models (DenseNet201, Xception, Mobilenet) followed by ML classification.
• Ensemble techniques were used to aggregate predictions for improved accuracy and robustness.
• Class weight updates and data augmentation strategies were employed to handle data imbalance.

Main Results:

• The proposed hybrid ML-DL model achieved high classification performance on HAM10000 and ISIC datasets.
• Achieved accuracies of 98.7% and 94.4%, with precision, recall, and F1-scores exceeding 95% for both datasets.
• Demonstrated significant improvements over existing methods in classification accuracy and generalization.

Conclusions:

• The combined ML-DL approach effectively classifies skin cancer, even with imbalanced data.
• This method offers a valuable tool for dermatologists, potentially improving early detection and patient care.
• The study highlights the potential of hybrid AI models in medical diagnostics.