Multispectral imaging and autofluorescence photobleaching combined in a multi-head neural network for skin cancer classification.

Introduction: The classification of hyperspectral images for skin cancer presents a significant challenge due to high data dimensionality and subtle differences between melanoma and non-melanoma tissues. This study investigates the efficacy of multi-target, multi-head neural network architectures for improving accuracy and precision-recall performance in hyperspectral melanoma classification. Methods: We use a proprietary multispectral dataset enriched with autofluorescence photobleaching tabular data, previously developed for skin lesion classification. Our method applies a multi-head architecture in which each head uses a different loss function, designed to optimize specific parts of the classification task. The network simultaneously learns from multiple data modalities, improving its ability to detect hidden features indicative of skin cancer. Final classifications are obtained by aggregating the outputs from all heads via simple averaging. Results: Results demonstrate significant improvements in classification accuracy and robustness compared to conventional single-head models. Our multi-head, multi-loss approach achieves the best performance on both evaluated data sources, with AUC-PR scores of 0.850 ± 0.032 and 0.822 ± 0.022 for the proprietary and ISIC datasets, respectively. Discussion: These findings indicate that multi-head architectures with specialized loss functions offer a powerful means of enhancing hyperspectral image classification, particularly for skin cancer detection, and provide a promising direction for future research and clinical applications.