Multiclass skin lesion localization and classification using deep learning based features fusion and selection...

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Smart healthcare leverages AI and IoT for efficient, individualized medical care.
• Teledermatology, a key application, uses telecommunications for remote skin lesion analysis.
• Current computational methods for skin cancer classification face challenges like low contrast and data imbalance.

Purpose of the Study:

• To propose a unified deep learning model for accurate skin lesion segmentation and classification.
• To enhance the efficiency and reliability of computer-aided diagnosis (CAD) in dermatology.
• To overcome limitations in existing methods for skin cancer detection.

Main Methods:

• Pre-processing dermoscopic images with a contrast enhancement technique.
• Utilizing a custom 26-layered convolutional neural network (CNN) for lesion segmentation.
• Employing transfer learning with modified pre-trained CNNs (Xception, ResNet-50, ResNet-101, VGG16) for classification.
• Fusing deep features and applying feature selection for final classification using a multi-class support vector machine (MC-SVM).

Main Results:

• The proposed model achieved high accuracy across multiple datasets: HAM10000 (98.57%), ISIC2018 (98.62%), ISIC2019 (93.47%), and PH2 (98.98%).
• Performance metrics demonstrate superior results compared to previous state-of-the-art methods.
• The approach effectively handles challenges such as low contrast and feature extraction.

Conclusions:

• The developed skin lesion detection and classification approach demonstrates superior performance.
• The model offers enhanced accuracy and quantitative evaluation for improved diagnostic capabilities.
• This deep learning framework represents a significant advancement in smart healthcare for dermatology.