Abstract
Purpose :
This study describes the application of deep learning methodology to identify the cornea and lens boundaries in Scheimpflug images of cataract patients for use in femtosecond laser-assisted cataract surgery (FLACS).
Methods :
Deidentified Scheimpflug images (973 eyes) were obtained, including 222 from LENSAR Laser System (LLS) surgeries performed in the commercial setting and the rest by LENSAR’s ALLY® device. The eyes imaged by ALLY® consisted of 628 from commercial surgeries, 64 from a clinical study environment, 31 non-cataractous eyes from pre-commercial clinical data collection, and 28 ex vivo scans obtained in the laboratory setting (9 human cadaver eyes and 19 porcine eyes). Images had the boundaries delineating the cornea and lens surfaces manually annotated. Aggressive data augmentation was performed during training, including illumination changes, noise addition, affine transforms, and histogram stretching. A deep convolutional neural network based on the U-Net architecture – with residual connections utilized within each block – was trained to identify pixels belonging to the anterior and posterior surfaces of the lens capsule and cornea. RANSAC circle detection is applied to resultant edge maps for the final output. During training, the RANSAC routine also applies an extra penalty to incorrect high-scoring curves on top of the cross-entropy loss function.
Results :
Twofold cross validation assessed the method on the 692 cataractous eyes imaged by ALLY®. Performance was measured on a per-eye basis in terms of 3D anatomical reconstruction leveraging known geometric constraints and outlier removal. There were zero failures to reconstruct either cornea surface or the lens anterior surface, and 5 failures to reconstruct the lens posterior surface, including 3 for which the surface was legitimately hidden due to cataract density. Successful reconstructions were screened for visible surface detection errors by comparing geometric quantities with those from the hand-labeled 3D reconstructions. A single case was found for the anterior lens surface while 8 were found for the posterior surface. Thus, the output produced a usable 3D lens reconstruction in 99.7% of cases and one free of discernible error in 98.8% of cases.
Conclusions :
Modern deep learning technology is capable of identifying the lens capsule even in the presence of advanced cataracts with a high degree of repeatability and accuracy.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.