Abstract
Purpose :
This study aims to create a novel method for accurately determining intraocular lenses (IOLs) positions after cataract surgery. Cataract surgery, being one of the most common worldwide, has advanced significantly, expanding its patient applicability. A crucial issue has been the post-surgical dislocation of IOLs within the capsule, impacting visual function and necessitating new surgical approaches. Current methods like UBM and AS-OCT sometimes fail, especially with Casia2 AS-OCT by Tomey Corporation. Hence, we've developed an artificial intelligence model using AS-OCT images to locate IOLs more precisely and aim to compare its accuracy with existing methods.
Methods :
We developed a machine learning model for IOL localization using AS-OCT images from Tsukazaki Hospital, analyzing 100 images and comparing them with the Casia2 system. The model was trained using the hold-out method with a dataset comprising 679 images: 455 training images from 33 individuals, 112 validation images from 8 individuals, and 100 test images. Each IOL image was annotated using Adobe Photoshop®.
Image processing included resizing to 224x224 pixels, rotations up to 15 degrees, horizontal flipping, and brightness-contrast adjustments. The segmentation estimation model used was UNet with DenseNet121, trained over 100 epochs with the Adam optimizer and Binary Cross Entropy Loss and Jaccard loss. Intersection over Union(IoU) score and tilt were used for performance evaluation.In the analysis using the Casia2 system, 100 cases were used. The IoU was calculated by measuring the areas of red regions determined from the anterior and posterior surfaces of the lens.
Results :
Using the existing Casia2 system, 9 instances of IOL identification failure, 7 misidentifications, and 16 cases where IOLs weren't detected were recorded. The traditional method’s IoU was 0.698 with 12 cases of IoU 0. The AI model had only one misidentification, achieving an IoU of 0.891 with no IoU 0 cases and one at 0.61. The traditional method’s tilt ranged from -6.6 to 7.4 degrees, whereas the AI’s was -6.2 to 7.5 degrees, calculating tilt in all cases.
Conclusions :
The machine learning model for IOL localization shows higher recognition efficiency than the current CASIA2 system. With the rapid progression of IOL technology, ongoing research is essential, but the current model’s accuracy is reliable enough for clinical application.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.