Enhancing Early GI Disease Detection with Spectral Visualization and Deep Learning.

Area of Science:

• Medical Imaging
• Gastroenterology
• Computer Vision

Background:

• Gastrointestinal disease diagnosis via endoscopy is limited by conventional white light imaging (WLI).
• Detecting early mucosal abnormalities is challenging due to low contrast and sensitivity.
• Advanced imaging techniques like hyperspectral imaging (HSI) offer potential but require specialized hardware.

Purpose of the Study:

• To develop and validate a software-driven framework, Spectrum Aided Vision Enhancer (SAVE).
• To transform standard WLI into HSI and simulated narrow-band imaging (NBI) without hardware modifications.
• To enhance the diagnostic performance for gastrointestinal diseases.

Main Methods:

• Utilized spectral reconstruction techniques: Macbeth Color Checker calibration, principal component analysis (PCA), and multivariate polynomial regression.
• Achieved high fidelity with RMSE of 0.056 and SSIM >90%.
• Trained and validated deep learning models (ResNet, EfficientNet) on the Kvasir v2 dataset (6490 images).

Main Results:

• SAVE-enhanced imagery consistently outperformed raw WLI in precision, recall, and F1-score.
• EfficientNet-B2 and EfficientNetV2-B0 models achieved the highest classification accuracy for GI conditions.
• Significant diagnostic performance improvement was demonstrated without specialized imaging hardware.

Conclusions:

• SAVE is a transformative software solution for augmenting gastrointestinal diagnostics.
• It has the potential to significantly improve early detection rates of GI diseases.
• SAVE can streamline clinical workflows and broaden access to advanced imaging, especially in resource-limited settings.