Dual-stage segmentation and classification framework for skin lesion analysis using deep neural network.

Area of Science:

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

Background:

• Skin diseases pose a global health burden, impacting physical and psychosocial well-being.
• Early and accurate detection of skin lesions, particularly malignant ones, is crucial for effective treatment and patient outcomes.
• Challenges in skin lesion analysis include data imbalance, variability in lesion appearance, and low image contrast.

Purpose of the Study:

• To develop and evaluate a dual-stage deep learning framework for precise skin lesion segmentation and classification.
• To address the challenges of imbalanced datasets and lesion variability in automated skin disease detection.
• To assess the framework's performance on benchmark datasets and a novel, clinically realistic dataset.

Main Methods:

• A U-Net model with a VGG16 encoder was employed for precise instance segmentation of skin lesions.
• EfficientFormer and SwiftFormer networks were utilized for lesion classification, tested on balanced and imbalanced data.
• Experiments were conducted on HAM10000, ISIC 2018, and the ISIC 2024 SLICE-3D datasets, including fusion approaches for the latter.

Main Results:

• The EfficientFormerV2 model achieved 97.11% accuracy and a 97.14% F1-score on the balanced HAM10000 dataset.
• The segmentation model demonstrated high performance on ISIC 2018, with 97.59% accuracy and 94.24% Dice similarity.
• On the ISIC 2024 SLICE-3D dataset, image+tabular fusion achieved a competitive score, highlighting performance in imbalanced, realistic settings.

Conclusions:

• The dual-stage deep learning framework exhibits high accuracy and robustness for skin lesion segmentation and classification.
• The framework's adaptability to large-scale, non-dermoscopic data like SLICE-3D indicates its potential for real-world applications.
• This approach shows promise for improving skin cancer triage and teledermatology services.