Abstract
INTRODUCTION: Multispectral and hyperspectral imaging have emerged as powerful tools in medical diagnostics, particularly in cancer detection, due to their ability to capture rich spectral information beyond human vision. Traditional approaches for cancer detection rely on handcrafted features and conventional machine learning algorithms, which struggle with high-dimensional spectral data, noise interference, and domain adaptation challenges. Deep learning has recently been introduced to address these limitations, yet existing models often lack robust feature extraction, generalization capability, and effective domain adaptation strategies. METHODS: In this study, we propose a novel deep learning-based time series prediction framework for multispectral and hyperspectral medical imaging analysis. Our approach integrates multi-scale feature extraction, attention mechanisms, and domain adaptation strategies to improve lesion segmentation and disease classification. The model employs self-supervised learning to mitigate the scarcity of labeled medical data, enhancing generalization across different imaging modalities. Furthermore, a knowledge-guided regularization module is introduced to leverage prior medical knowledge, refining predictions and reducing false positives. RESULTS: Experimental results demonstrate that our framework outperforms state-of-the-art methods in spectral imaging-based cancer detection, achieving superior accuracy, robustness, and interpretability. DISCUSSION: The proposed approach provides a significant step toward AI-driven medical imaging solutions that effectively harness multispectral and hyperspectral data for enhanced diagnostic performance.