Abstract
Early esophageal squamous neoplasia is difficult to detect under conventional white light imaging due to subtle mucosal and vascular changes. This study utilizes a spectrum-aided vision enhancement (SAVE) framework to computationally reconstruct 401-band hyperspectral information from standard RGB endoscopy images. By integrating these virtual spectral features with a U-Net semantic segmentation model, the approach enhances lesion boundary delineation without requiring specialized hyperspectral hardware. Evaluation on 531 clinical images demonstrates that models using virtual spectral data achieve a mean intersection over union of 74.3%, outperforming conventional white light and narrow-band imaging baselines. Cross-center validation further confirms the generalizability of this model-agnostic enhancement across different endoscopy systems. These findings indicate that virtual hyperspectral reconstruction provides a feasible strategy for enriching diagnostic features in routine clinical workflows. This approach offers a scalable tool for improving the computer-aided detection of early-stage gastrointestinal malignancies.