Diagnostic Accuracy Of A Deep Learning Model For Automated Assessment Of Toric Intraocular Lens Rotation

To develop a deep learning (DL) model for automated assessment of toric intraocular lens (IOL) rotation from retroillumination images using scleral landmarks as a reference, thereby minimizing the influence of cyclorotation.

Two datasets of paired retroillumination images obtained 1 hour and 6 months after cataract surgery were manually graded for toric IOL rotation using scleral landmarks as reference. Dataset A included 125 eyes implanted with a Clareon toric IOL and dataset B included 120 eyes implanted with a Rayner EMV toric IOL. A deep learning framework combining a convolutional neural network (CNN) with additional visual computing algorithms was trained, validated, and tested. Model performance was evaluated in a binary classification task for detecting IOL rotation >2°, using decision thresholds of 2° and 5° for the predicted rotation. Performance metrics included sensitivity, specificity, and accuracy.

For the Clareon toric IOL, at a 2° threshold, sensitivity, specificity, and accuracy were 0.50, 0.56, and 0.56, respectively. At 5°, sensitivity remained 0.50, while specificity increased to 1.00, resulting in an accuracy of 0.94. For the Rayner EMV toric IOL, at 2°, sensitivity, specificity, and accuracy were 0.82, 0.83, and 0.82, respectively. At 5°, sensitivity decreased to 0.73, specificity increased to 1.00, and overall accuracy was 0.91.

The developed DL model demonstrated promising performance in the assessment of toric IOL rotation. Increasing the threshold from 2° to 5° substantially improved specificity and overall accuracy for both IOL types, at the expense of sensitivity. Superior performance was observed for the Rayner EMV IOL, likely due to enhanced visibility of the toric IOL markings.