Skin Cancer Segmentation and Feature Extraction of Dermatoscopy Images Based on Deep Learning U-Net

Skin cancer in Indonesia represents 5.9%-7.8<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> of the total cancer cases, early detection can increase the cure rate by up to 90%. Despite this, early detection can be complex and subjective, resulting in delays in diagnosis. Computer-Aided Diagnostic (CAD) systems are developed to improve diagnostic accuracy. However, due to the complex variations in the appearance of different lesions, automated diagnosis faces challenges in achieving consistent accuracy and avoiding false diagnoses. In this study, we propose a system that includes image preprocessing to enhance image quality, image segmentation using U-Net to separate lesions from the background, and feature extraction using the ABCD method, which analyzes asymmetry (A), border (B), color (C), and diameter (D) of skin lesions. The U-Net model, trained with 50 epochs and a learning rate 0.0001, achieved evaluation scores of 0.81 for Dice Similarity Coefficient, 0.25 for loss, and 0.68 for Intersection over Union on the test data. Lesions with high asymmetry and irregular perimeters are often associated with malignant skin cancers such as Actinic Keratosis, Basal Cell Carcinoma, and Melanoma. In addition, the high red intensity of the lesion may indicate inflammation or increased vascularization, which are critical factors in assessing and diagnosing skin lesions.