Correcting Axial Length-Related Magnification Errors in Optical Coherence Tomography Angiography Improves Machine Learning Classification Performance in High Myopia.

BACKGROUND

To determine whether axial length-related magnification correction in optical coherence tomography angiography (OCTA) improves machine learning classification of high myopia and resolves inconsistencies in the values and discriminatory power of OCTA microvascular metrics.

METHODS

We analysed retinal OCTA images from a prior Hong Kong Polytechnic University study. Fourteen superficial vascular plexus metrics were extracted before and after magnification correction and used to train random forest (RF) models to classify high versus non-high myopia. Outcomes included classifier performance (the area under the receiver operating characteristic curve (AUC), sensitivity, specificity and feature importance) and changes in OCTA metrics between groups before and after correction.

RESULTS

Image magnification correction significantly improved the classification of high versus non-high myopia. When all 14 features were used to train the model, the AUC increased from 0.77 (fair) to 0.88 (good), and when restricted to the top five features (fractal dimension (Df), vessel length density (VLD), branchpoint density (BD), parafoveal rim VLD, and vessel area density (VAD)), similarly, from 0.79 to 0.88.

CONCLUSIONS

OCTA image magnification correction improves RF classification of high myopia and clarifies inconsistencies across studies. Underreported metrics, fractal dimension, parafoveal rim VLD and BD, emerged as key discriminators. Magnification correction with ML may improve characterisation of myopia-related microvascular changes and enhance diagnostic precision, warranting further study.