TRANSFER LEARNING-BASED DEEP ARCHITECTURE FOR VITILIGO DETECTION WITH AN IMPROVED SEGNET MODEL

As a chronic condition resulting in pigmentation loss, vitiligo requires early and precise detection for effective medical intervention. This work presents a dual-stage deep learning system that combines improved preprocessing, segmentation, and classification specifically designed for vitiligo detection, in contrast to previous research that depends on traditional CNN architectures or single-stage pipelines. An innovative architectural development that greatly raises the accuracy of multi-scale feature extraction and classification is the incorporation of FSMSDSCM and WESELU into the IGNT-TL model. A new deep learning approach for vitiligo detection is proposed in this paper, leveraging Transfer Learning with an Improved SegNet (ImSNT) for segmentation and an optimized GoogleNet for classification. Initially, in the preprocessing stage, an Improved Anisotropic Filtering (IAF) technique is introduced, incorporating the Inverse Laplace of Gaussian (ILOG) transformation with a Wiener filter in Anisotropic Diffusion Filtering (ADF) to effectively remove noise while preserving essential details like edges and textures. The segmentation stage employs an ImSNT model, which captures multi-scale spatial information, leading to precise segmentation of vitiligo-affected skin regions. In the feature extraction stage, four crucial feature types are derived from segmented images: texture features using Gray-Level Co-occurrence Matrix (GLCM), disease area features, pixel-based features, and color features using RGB histograms. These extracted features form a comprehensive dataset for classification. In the detection stage, an Improved GoogleNet model with Transfer Learning (IGNT-TL) is introduced, utilizing a Feature Slicing-Multisize Depthwise Separable Convolution Module (FSMSDSCM) for multi-scale feature extraction and an optimized Weighted ELiSH Scaled Exponential Linear Unit (WESELU) activation function for better gradient propagation and classification performance. The proposed two-stage transfer learning approach enhances model accuracy while reducing overfitting. This research significantly contributes to automated vitiligo detection, providing a robust, efficient, and accurate DL framework for medical image analysis. The highest performance was recorded by the IGNT-TL strategy, achieving 0.964 accuracy, 0.980 precision, and 0.968 [Formula: see text]-measure. In addition, the IGNT-TL model exhibited the least error rates, with an FNR of 0.044, FPR of 0.025, and FDR of 0.020.