MSCDNet-based multi-class classification of skin cancer using dermoscopy images

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer poses a significant health risk, with early detection dramatically improving patient outcomes.
• Deep learning models are increasingly utilized for skin cancer detection.
• This research focuses on classifying multiple skin cancer types, including melanoma, basal cell carcinoma, melanocytic nevi, and benign keratosis.

Purpose of the Study:

• To introduce a novel deep learning-based network model for the automatic classification of multiple skin cancer types.
• To develop an efficient semantic segmentation model for skin lesions.
• To enhance the accuracy and efficiency of skin cancer diagnosis.

Main Methods:

• A Multi-class Skin Cancer Detection Network (MSCD-Net) was developed.
• An efficient semantic segmentation deep learning model, DenseUNet, was proposed for skin lesion segmentation.
• Feature selection was performed using the Binary Dragonfly Algorithm (BDA), followed by SqueezeNet-based classification.

Main Results:

• The DenseUNet model, leveraging DenseNet connections and UNet architecture, achieved superior segmentation results by utilizing low-level features.
• The proposed MSCD-Net model demonstrated enhanced effectiveness and efficiency compared to previous methods on the ISIC 2019 dataset.
• Performance was rigorously evaluated using the standard ISIC 2019 dataset.

Conclusions:

• The developed MSCD-Net model offers a promising advancement in automated multi-class skin cancer detection.
• The integration of DenseUNet for segmentation and BDA for feature selection contributes to improved diagnostic accuracy.
• The findings suggest that this deep learning approach can significantly aid in the early and efficient identification of various skin cancers.