Two-step hierarchical binary classification of cancerous skin lesions using transfer learning and the random forest...

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Skin lesion classification is vital for early disease detection and diagnosis.
• Computer-aided diagnostic tools improve patient outcomes, especially in underserved areas.
• Current deep learning models struggle with limited and imbalanced skin lesion datasets.

Purpose of the Study:

• To develop a hybrid machine and deep learning model for improved skin lesion classification.
• To address challenges posed by data imbalance and limited dataset sizes in skin disease detection.
• To enhance the efficiency and accuracy of automated skin lesion diagnosis.

Main Methods:

• A two-step hierarchical binary classification strategy was implemented.
• DenseNet121 (DNET) was utilized as a feature extractor.
• A random forest (RF) classifier was employed in conjunction with deep learning (DL).
• Experiments were conducted on the International Skin Imaging Collaboration (ISIC 2017) dataset.

Main Results:

• The hybrid hierarchical approach achieved a balanced multiclass accuracy (BMA) of 91.07%.
• This outperformed a pure deep-learning model (end-to-end DNET) which achieved 88.66% BMA.
• The random forest ensemble demonstrated superior efficiency compared to other machine learning classifiers.
• The model significantly reduced computational time.

Conclusions:

• The proposed hybrid hierarchical model effectively handles large class imbalances in skin lesion datasets.
• This approach enhances classification performance and efficiency for real-world applications.
• The findings suggest a promising direction for improving computer-aided diagnosis in dermatology.