Machine learning-based diagnosis of melanoma using macro images

Area of Science:

• Medical Imaging
• Computational Dermatology
• Artificial Intelligence in Medicine

Background:

• Melanoma diagnosis relies heavily on dermatoscopic images, but a gap exists for general-purpose cameras.
• Non-uniform illumination in macro images significantly hinders accurate information extraction for skin cancer detection.

Purpose of the Study:

• To develop a computer-aided decision support system for melanoma diagnosis using macro images from general cameras.
• To address challenges posed by non-uniform illumination and improve lesion segmentation and classification accuracy.

Main Methods:

• Multistage illumination compensation for creating a smooth illumination surface.
• Multimode segmentation for infected region extraction.
• Feature extraction (geometry, photometry, border, texture) with information theory for redundancy reduction.
• Classification using Support Vector Machine, Random Forest, Neural Network, and Naive Bayesian classifiers.

Main Results:

• The proposed model demonstrates improved performance in distinguishing benign from malignant melanoma.
• Experimental outcomes supported by hypothesis testing and boxplot representation validate classification accuracy.
• The system effectively mitigates issues caused by non-uniform illumination.

Conclusions:

• The developed computer-aided system offers a significant advancement for melanoma diagnosis using readily available camera technology.
• The multistage illumination compensation and multimode segmentation methods are crucial for enhancing diagnostic accuracy.
• The study highlights the potential of AI in improving skin cancer detection from non-specialized imaging devices.