Glaucoma Detection through a Novel Hyperspectral Imaging Band Selection and Vision Transformer Integration

Area of Science:

• Ophthalmology
• Medical Imaging
• Computational Biology

Background:

• Glaucoma diagnosis traditionally uses fundus images, focusing on optic cup-to-disc ratio and retinal abnormalities.
• Hyperspectral imaging offers a novel approach by detecting changes in retinal vessel oxygen saturation.

Purpose of the Study:

• To explore the diagnostic potential of hyperspectral imaging for glaucoma.
• To introduce a new technique for converting digital fundus images into hyperspectral representations for detailed spectral analysis.

Main Methods:

• Digital fundus images were converted to hyperspectral representations.
• Spectral analysis identified key spectral regions for image reconstruction.
• Vision Transformer (ViT) algorithm was used for classification across selected spectral bands.
• A dataset of 1291 fundus images was utilized.

Main Results:

• The 610-780 nm spectral band demonstrated superior performance, achieving approximately 0.9007 in accuracy, precision, recall, F1-score, and AUC-ROC.
• The RGB model also showed strong classification results.
• Other spectral bands exhibited lower recall and overall performance metrics.

Conclusions:

• Hyperspectral imaging, combined with Vision Transformer (ViT) algorithms, shows significant potential to improve glaucoma diagnosis.
• This approach offers a more comprehensive analysis compared to traditional fundus image methods.
• The findings suggest a transformation in glaucoma diagnostics through integrated hyperspectral imaging and computational techniques.