Refined Residual Deep Convolutional Network for Skin Lesion Classification.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer is the most common human cancer, often diagnosed via clinical screening and dermoscopy.
• Automated skin lesion classification is challenging due to visual similarities between melanoma and benign lesions.

Purpose of the Study:

• To propose a novel residual deep convolutional neural network (RDCNN) for improved skin lesion diagnosis.
• To evaluate the performance of the proposed RDCNN across multiple benchmark skin cancer datasets.

Main Methods:

• The RDCNN was trained and tested on six datasets: PH2, DermIS and Quest, MED-NODE, ISIC2016, ISIC2017, and ISIC2018.
• Three experiments were conducted: training on original images, testing on segmented images, and transfer learning using a pre-trained model.

Main Results:

• The proposed RDCNN demonstrated high performance in classifying skin lesions.
• The RDCNN model achieved superior results compared to existing deep convolutional networks.

Conclusions:

• The developed RDCNN offers a promising approach for automated skin lesion diagnosis.
• The RDCNN's effectiveness is validated across diverse datasets and experimental conditions.