Generative-augmented deep learning framework for scarce-data skin lesion image processing.

PURPOSE: Skin conditions are health conditions that affect millions of people across the globe and are a major health challenge in the world. The existing skin lesion analysis models have poor performance under scarce and imbalanced data conditions, which leads to reduced generalization across diverse lesion types. Most prior models lack adaptability to real-world Internet of Things (IoT)-based acquisition settings and heterogeneous imaging environments. MATERIALS AND METHODS: The proposed system combines lesion image acquisition based on IoT with various public datasets to enhance the variability of data to overcome these problems. High-level preprocessing improves lesion visibility and eliminates artifacts, whereas controlled augmentation reduces the imbalance. To capture fine-grained lesion details, an integrated U-shaped Attention Network (UAN) with Convolutional Neural Network and Vision Transformer (CNN-ViT)-based feature extraction is used. A hybrid Kolmogorov-Arnold Network and Binary Spiking Neural Network (KAN-BSNN) classifier is used to enhance robustness and decision boundaries and then optimized using the Dream-Coati Optimization (DCO) algorithm. The Explainable AI(XAI)methods such as Local Interpretable Model-agnostic Explanations (LIME) and counterfactual analysis ensure transparent, clinically interpretable and reliable diagnostic decision support. RESULTS: The proposed approach achieves 99.27% accuracy, 99.01% precision, 98.59% recall, 98.27% F-measure and 99.02% of specificity. CONCLUSIONS: Furthermore, the suggested model indicates better classification stability and a decreased error rate with less computational complexity.