Skin Lesion Diagnosis Through Deep Learning and Hybrid Texture Feature Augmentation

Irpan Adiputra Pardosi (Department of Informatics Engineering, Universitas Mikroskil, Medan, North Sumatera, Indonesia)
Roni Yunis (Department of Information Systems, Universitas Mikroskil, Medan, North Sumatera, Indonesia)
Arwin Halim (Department of Informatics Engineering, Universitas Mikroskil, Medan, North Sumatera, Indonesia)

Teknika
Vol. 14 No. 2 (2025): July 2025

Skin cancer is a leading cause of cancer-related deaths globally, with melanoma being the most lethal subtype. Early detection remains critical for improving patient outcomes. However, dermoscopic image analysis faces challenges due to inter-class similarity between malignant melanoma and benign nevi. This study proposes a robust framework for optimizing Gray-Level Co-occurrence Matrix (GLCM) and Local Binary Pattern (LBP) parameters using the ISIC 2023 dataset. The framework integrates handcrafted features with Convolutional Neural Networks (CNNs) to enhance classification accuracy. Key contributions include: Automated parameter tuning for GLCM and LBP using grid search and cross-validation; A hybrid model combining EfficientNet-B3 with full handcrafted features; Comprehensive evaluation on the ISIC 2023 dataset (10,015 images). Results demonstrate that the hybrid model (Scenario 2) achieves 93.7% accuracy and 92.8% F1-score, outperforming the standalone CNN model (Scenario 1) by 3.5%. The proposed framework reduces false positives by 15% compared to dermatologist assessments, highlighting its potential for clinical decision support. Future work will explore advanced architectures like EfficientNet-B4 and integration of external factors such as lesion location.