José Carlos Fernández-Álvarez1, Iván Hernández-López2, Pedro Pablo Cruz-Cobas3, Taimí Cárdenas-Díaz2, Alfo José Batista-Leyva4. 1. Instituto Superior de Tecnologías y Ciencias Aplicadas, Universidad de la Habana, la Habana, Cuba. 2. Instituto Cubano de Oftalmología Ramón Pando Ferrer, la Habana, Cuba. 3. Facultad de Matemática, Universidad de la Habana, la Habana, Cuba. 4. Instituto Superior de Tecnologías y Ciencias Aplicadas, Universidad de la Habana, la Habana, Cuba. Electronic address: abatista@instec.cu.
Abstract
PURPOSE: To determine the capabilities of a multilayer perceptron (MLP) for calculating the power of an intraocular lens (IOL) to be implanted and in achieving a given postoperative stable refraction. SETTING: Cuban Institute of Ophthalmology, Havana, Cuba. DESIGN: Retrospective review. METHODS: The study comprised data of patients who had uneventful phacoemulsification cataract surgery with implantation of a biconvex acrylic foldable IOL (type RYCF, model Ocuflex) in the capsular bag over 6 years. Exclusion criteria were previous intraocular or refractive corneal surgery, any corneal disease, pathological or complicated cataracts, intraoperative complications, preoperative astigmatism beyond 3.0 diopters (D), postoperative corrected distance visual acuity worse than 20/40, missing postoperative refractive information, eyes with an axial length (AL) shorter than 19.36 mm, eyes with an AL longer than 27.0 mm, average corneal keratometry (K) power lower than 36.0 D or higher than 50.9 D, and refractive surprises greater than ±3.0 D. The data were used to train an MLP to predict the value of the IOL power required for attaining a given postoperative refraction. Using AL, K value, and predicted and real postoperative refraction as input data, the output of the MLP was the IOL power. RESULTS: The study comprised 15 728 eyes of 15 728 patients. The trained neural networks predicted the value of the implanted IOL with an error less than 0.5 D in more than 95% of patients, even for a case in which a surgeon was not included in the training process. CONCLUSIONS: The accuracy attained by the trained MLP is high, indicating the feasibility of a prospective study leading to a new method of predicting the IOL power in refractive surgery with an error lower than the current prediction methods.
PURPOSE: To determine the capabilities of a multilayer perceptron (MLP) for calculating the power of an intraocular lens (IOL) to be implanted and in achieving a given postoperative stable refraction. SETTING: Cuban Institute of Ophthalmology, Havana, Cuba. DESIGN: Retrospective review. METHODS: The study comprised data of patients who had uneventful phacoemulsification cataract surgery with implantation of a biconvex acrylic foldable IOL (type RYCF, model Ocuflex) in the capsular bag over 6 years. Exclusion criteria were previous intraocular or refractive corneal surgery, any corneal disease, pathological or complicated cataracts, intraoperative complications, preoperative astigmatism beyond 3.0 diopters (D), postoperative corrected distance visual acuity worse than 20/40, missing postoperative refractive information, eyes with an axial length (AL) shorter than 19.36 mm, eyes with an AL longer than 27.0 mm, average corneal keratometry (K) power lower than 36.0 D or higher than 50.9 D, and refractive surprises greater than ±3.0 D. The data were used to train an MLP to predict the value of the IOL power required for attaining a given postoperative refraction. Using AL, K value, and predicted and real postoperative refraction as input data, the output of the MLP was the IOL power. RESULTS: The study comprised 15 728 eyes of 15 728 patients. The trained neural networks predicted the value of the implanted IOL with an error less than 0.5 D in more than 95% of patients, even for a case in which a surgeon was not included in the training process. CONCLUSIONS: The accuracy attained by the trained MLP is high, indicating the feasibility of a prospective study leading to a new method of predicting the IOL power in refractive surgery with an error lower than the current prediction methods.