Stereo-Vision-Inspired Spectral Disparity Learning for Choroidal Tumor Thickness Prediction.

Purpose: To investigate whether choroidal tumor thickness can be estimated from fundus photographs using a stereo-vision-inspired deep learning approach. Methods: A deep learning framework was developed to independently extract features from red and green color channels of ultra-widefield fundus photographs and integrate them using a spectral disparity module (SDM). The SDM explicitly captures inter-channel correspondence and disparity, mimicking the principles of human binocular depth perception. The model was trained and evaluated on 337 patients with choroidal tumors, using ultrasound-measured thickness as ground truth. Performance was assessed using regression metrics and risk classification based on the Collaborative Ocular Melanoma Study criteria. Results: The proposed stereo-fusion model achieved strong agreement with the ground-truth ultrasound measurements (root mean squared error = 1.07 mm, mean absolute error = 0.74 mm, R² = 0.85), outperforming single-channel and conventional color image models. Thickness-based risk stratification yielded an accuracy of 90.5%, a sensitivity of 87.7%, and a specificity of 92.9%. Conclusion: Choroidal tumor thickness can be accurately estimated from fundus photographs by leveraging wavelength-dependent spectral disparities. This approach enables noncontact thickness estimation and supports scalable screening and triage of choroidal tumors. Translational Relevance: By leveraging wavelength-dependent spectral disparities with deep learning, this work translates fundus photography into a noncontact tool for estimating choroidal tumor thickness.