Melanoma Diagnosis Using Deep Learning and Fuzzy Logic.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Melanoma is a dangerous skin cancer requiring early detection.
• Distinguishing benign from malignant lesions can be challenging.
• Deep convolutional neural networks (DCNNs) show promise in medical image analysis.

Purpose of the Study:

• To develop a faster and more precise deep learning algorithm for melanoma detection.
• To improve upon conventional Convolutional Neural Network (CNN) performance in melanoma diagnosis.
• To integrate advanced image processing techniques for enhanced feature extraction.

Main Methods:

• Implementation of a deep learning-based 'You Only Look Once (YOLO)' algorithm using DCNNs.
• Prediction of bounding boxes and class confidence scores for detected lesions.
• Inclusion of two-phase segmentation using graph theory (minimal spanning tree) and L-type fuzzy number approximations for precise area calculation.

Main Results:

• The YOLO algorithm achieved high performance on multiple datasets.
• Jac scores of 79.84% (ISIC 2019), 86.99% (ISBI 2017), and 88.64% (PH2) were obtained.
• The proposed method demonstrated superior results compared to existing approaches in most cases.

Conclusions:

• The developed YOLO-based deep learning model offers a computationally effective solution for melanoma detection.
• The integration of novel segmentation and feature extraction methods enhances diagnostic accuracy.
• This approach holds potential for earlier and more reliable diagnosis of melanoma.