Irt Simulated Hyperspectral Deep Learning For Accurate Melanoma Detection

Abstract

Early detection of melanoma is crucial for improving survival rates, yet current diagnostic methods often rely on subjective analysis and limited imaging modalities. This study proposes a novel IRT‑simulated hyperspectral deep learning framework for skin cancer classification using the Kaggle Skin Cancer: Malignant vs. Benign dataset. Initially, RGB dermoscopic images are synthetically transformed into simulated Infrared Thermography (IRT) thermal maps, generating steady‑state and rewarming phases to reveal subsurface lesion characteristics. These are further converted into hyperspectral‑like representations through spectral decomposition, producing rich spectral information for lesion differentiation. Pre‑processing involves background removal, morphological hair filtering, and hue‑gamma correction to enhance lesion visibility. Spectral‑spatial features are then extracted and refined using the Whale Optimization Algorithm (WOA) to retain the most relevant and non‑redundant features. Classification is performed using NASceptionNet, a hybrid model combining NASNet and XceptionNet, enabling robust and accurate lesion classification. The framework is evaluated using multiple performance metrics, demonstrating its potential as a non‑invasive, accurate, and efficient screening tool for early melanoma detection, bridging the gap between dermoscopic imaging and advanced spectral analysis. The proposed model demonstrated remarkable improvement after IRT conversion. Before IRT conversion, the model achieved 98.04% accuracy, 97.88% precision, 98.21% recall, and 98.04% F1-score, with an MCC of 96.09. After IRT conversion, performance increased to 99.12% accuracy, 99.05% precision, 99.18% recall, and 99.11% F1-score, with an MCC of 98.23. These results indicate a consistent 1–2% performance boost across all metrics, confirming the effectiveness of IRT conversion in enhancing diagnostic accuracy.