Automatic Detection of Malignant Melanoma using Macroscopic Images.

Area of Science:

• Dermatology
• Medical Imaging
• Computer Vision

Background:

• Distinguishing benign from malignant pigmented skin lesions is crucial for early melanoma detection.
• Existing computerized methods often require specialized equipment or controlled imaging conditions.

Purpose of the Study:

• To develop a novel, user-friendly computerized procedure for detecting malignant melanoma from benign pigmented lesions using conventional digital cameras.
• To enhance image processing techniques for improved lesion analysis under unconstrained imaging conditions.

Main Methods:

• A new procedure was developed for analyzing macroscopic images of pigmented skin lesions taken with standard digital cameras.
• Image processing techniques were introduced to correct for non-uniform illumination, hair, and glow artifacts.
• A threshold-based segmentation algorithm was employed, followed by feature extraction (187 features) representing lesion characteristics.
• Principal Component Analysis (PCA) was used for feature reduction, and a Support Vector Machine (SVM) classifier was utilized for lesion classification.

Main Results:

• The proposed method achieved high accuracy in detecting lesion areas, comparable to dermatological diagnoses.
• Feature reduction using PCA identified 13 key features that yielded superior classification results compared to using all extracted features.
• The classification evaluation demonstrated an overall accuracy of 82.2%, with a sensitivity of 77% and specificity of 86.93%.

Conclusions:

• The developed computerized procedure effectively distinguishes between benign and malignant pigmented skin lesions using conventional digital cameras.
• The method's minimal constraints, ease of use, and high accuracy offer a valuable tool for early-stage malignant lesion detection.
• This technique has the potential to assist dermatologists and non-experts in identifying potentially malignant lesions during primary screening.