TriDermCancerNet: A hybrid deep learning framework for skin cancer classification

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated skin cancer diagnosis from dermoscopic images is challenging due to poor image quality, similar lesion appearances, and imbalanced datasets.
• Existing methods struggle with variability and interclass similarities in skin lesion classification.

Purpose of the Study:

• To propose a novel Tri Model Dermatology Cancer Neural Network (TriDermCancerNet) for accurate skin cancer classification.
• To address challenges including dataset variability, interclass similarity, and model explainability in skin cancer diagnosis.

Main Methods:

• Utilized International Skin Imaging Collaboration 2018 and 2019 datasets.
• Applied contrast enhancement and data augmentation for image preprocessing and dataset balancing.
• Developed TriDermCancerNet integrating Inception, Inverted Bottleneck Residual, and Dense modules in parallel, with feature fusion via depth concatenation.
• Optimized hyperparameters using Bayesian optimization.

Main Results:

• The fused TriDermCancerNet achieved 98.6% accuracy and 1.0 AUC on the ISIC 2018 dataset.
• Achieved 99.7% accuracy and 1.0 AUC on the ISIC 2019 dataset.
• Statistical testing confirmed the superiority of the fused model over individual branches (p < 0.05).

Conclusions:

• The proposed TriDermCancerNet offers a precise and robust framework for skin cancer classification.
• This hybrid deep learning approach serves as a valuable diagnostic aid for clinicians in early skin cancer detection.