Automatic diagnosis of melanoma using machine learning methods on a spectroscopic system

Area of Science:

• Dermatology and Medical Imaging
• Biomedical Engineering
• Computational Pathology

Background:

• Early melanoma diagnosis is critical for reducing mortality but faces challenges in accuracy and subjectivity.
• Distinguishing melanoma from benign moles requires complex laboratory tests.
• Current diagnostic methods can be subjective and time-consuming for dermatologists.

Purpose of the Study:

• To develop an automated, non-subjective approach for melanoma diagnosis using quantitative measures.
• To improve the accuracy and efficiency of early skin cancer detection.
• To provide dermatologists with a reliable tool for objective diagnostic decision-making.

Main Methods:

• Spectroscopic device combining polarized and un-polarized light for image acquisition.
• Image processing including noise reduction and normalization.
• Feature extraction based on statistical measurements of pixel intensities.
• Classification using artificial neural network, naïve bayes, and k-nearest neighbour models.

Main Results:

• The naïve bayes classifier achieved 89% accuracy, 89% sensitivity, and 89% specificity.
• The developed approach was integrated into a desktop application for real-time diagnosis.
• The system demonstrated effective characterization of melanoma patterns through statistical feature analysis.

Conclusions:

• The approach utilizes advanced spectroscopic techniques to analyze multilayered skin characteristics.
• Image segmentation is not required, as the system directly probes lesion spots.
• The automated diagnostic tool offers a non-subjective second opinion, enhancing clinical decision support for melanoma.