Advanced Deep Learning Models for Melanoma Diagnosis in Computer-Aided Skin Cancer Detection.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Melanoma is the most lethal skin cancer, often diagnosed late due to limitations in visual examination.
• Early detection and prompt treatment are vital for limiting melanoma's severity.

Purpose of the Study:

• To develop efficient computer-aided diagnosis (CAD) approaches for melanoma detection.
• To enhance preprocessing, segmentation, and classification stages for improved diagnostic accuracy.

Main Methods:

• A hybrid method combining morphological operations and deep neural networks for image preprocessing (hairline removal, contrast enhancement).
• A deep learning-based segmentation network to isolate lesion regions.
• A deep neural network for classifying melanoma versus benign lesions using the ISIC 2020 dataset.

Main Results:

• The classification model incorporating preprocessing and segmentation achieved a 93.40% accuracy rate.
• The model demonstrated an efficient test time of 1.3 seconds per image.
• Comparison showed superior performance when using cleaned and segmented images versus raw images.

Conclusions:

• Automated CAD models significantly improve early skin cancer detection accuracy.
• Preprocessing and segmentation are critical steps for enhancing melanoma classification performance.
• The proposed methods offer a promising tool for assisting dermatologists in melanoma diagnosis.