Texture-based deep learning for accurate classification of pigmented skin lesions using autoencoder based hair filtered balanced imageset

Texture analysis plays a crucial role in image processing, particularly for classification tasks in medical imaging. In dermatology, accurate classification of pigmented skin lesion images is essential for early detection and diagnosis of skin cancers, notably melanoma—the most aggressive form. This study presents a texture-based deep learning framework for the classification of pigmented skin lesions. Gray Level Co-occurrence Matrix (GLCM) features, including contrast, homogeneity, and entropy, are extracted to characterize lesion texture in terms of roughness, regularity, and randomness. These features are subsequently used as inputs to various machine learning and deep learning classifiers. To address image artifacts, hair removal is performed using a two-stage filtering process combining image processing techniques and a convolutional autoencoder. Furthermore, image augmentation is applied to balance the dataset and improve model generalization. Experimental results demonstrate that the proposed approach achieves notable accuracy improvements when integrated with deep learning architectures, underscoring its potential for reliable skin lesion classification in clinical settings. • Autoencoder-based filtering removes hair artifacts from skin lesion images • Balanced dataset improves classification fairness across lesion categories • Texture features boost deep learning accuracy for lesion classification • Achieved high accuracy in multi-class pigmented skin lesion recognition • Method shows potential for robust, real-time dermatological diagnostics