An Optimal Deep Hybrid Framework with Selective Kernel U-Net for Skin Lesion Detection and Classification.

Area of Science:

• Dermatology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Early skin cancer detection, especially malignant melanoma, is crucial for reducing mortality.
• Automated analysis of dermoscopic images offers potential to enhance clinical diagnosis over manual inspection.
• Challenges in automated analysis include image noise, low contrast, lesion variability, and redundant feature representation.

Purpose of the Study:

• To develop an optimal deep hybrid framework for robust and efficient skin lesion detection and classification.
• To integrate advanced preprocessing, precise segmentation, optimal feature selection, and accurate classification for improved diagnostic performance.
• To address limitations of existing methods by proposing a novel framework for skin cancer analysis.

Main Methods:

• Image preprocessing using Contrast Limited Adaptive Histogram Equalization (CLAHE) and Wiener filtering for contrast enhancement and noise reduction.
• Lesion segmentation via Selective Kernel U-Net (SK-UNet) to capture multi-scale spatial information.
• Feature extraction and optimization using the Fossa Optimization Algorithm (FOA) followed by classification with a hybrid 1D-CNN-GRU model.

Main Results:

• The proposed framework achieved high classification accuracies: 97.6% on the ISIC dataset and 95.6% on the DermMNIST dataset.
• The system demonstrated superior performance compared to several existing methods in skin lesion classification.
• The optimized feature selection effectively reduced redundancy and improved classification precision.

Conclusions:

• The developed deep hybrid framework provides a reliable, accurate, and scalable solution for skin cancer diagnosis.
• The system shows significant potential in assisting clinical decision-making and facilitating early detection of skin cancer.
• The integration of advanced techniques in preprocessing, segmentation, feature selection, and classification leads to improved diagnostic outcomes.