Flag-Net: Classification of Skin Lesions with a Hybrid Deep Learning Approach Based on Fractals and Lacunarity

OpenAlex Turkish

In recent years, artificial intelligence-based methods, particularly deep learning, have achieved significant success in medical image analysis. This study proposes FLAG-Net, a hybrid deep learning model designed to overcome traditional CNN limitations by enhancing structural sensitivity through fractal dimension and lacunarity-based texture analysis. FLAG-Net enriches multilevel convolutional features with an attention mechanism and integrates morphological and fractal structure maps to improve classification performance. The model was evaluated on the HAM10000 and ISIC 2019 skin lesion datasets, achieving accuracies of 98.54% and 98.72%, respectively—outperforming well-known architectures such as InceptionV3, EfficientNet, VGG19, and ResNet50. Ablation studies were performed to analyze the contribution of key components individually, confirming that the attention mechanism, multilevel feature fusion, and fractal/lacunarity maps significantly enhance classification results. Overall, FLAG-Net not only achieves high accuracy but also strengthens decision-making by effectively capturing complex texture patterns. The findings highlight FLAG-Net’s potential as a reliable and generalizable model with strong clinical applicability in medical image classification.