Search

RU-2861597-C1 - METHOD FOR CALCULATING OPTICAL POWER OF INTRAOCULAR LENS BASED ON ARTIFICIAL NEURAL NETWORK MODELS WITH PRELIMINARY CLASSIFICATION OF PATIENT GROUPS

RU2861597C1RU 2861597 C1RU2861597 C1RU 2861597C1RU-2861597-C1

Abstract

FIELD: ophthalmology. SUBSTANCE: invention can be used when calculating the optical power of an intraocular lens (IOL) using mathematical models obtained as a result of deep learning of artificial neural networks (ANN models). Four ANN models are used, the first of which acts as a classifier and directs a patient, based on his preoperative examination data: refraction of the strong meridian, refraction of the weak meridian, optical eye length, anterior chamber depth, lens thickness, a-constant of the model of the implanted IOL, to one of three calculating ANN models, each of which has its own architecture and calculates the IOL optical power only for the group of patients obtained during classification by the first ANN model. The first group consists of patients for whom the calculated IOL values do not exceed the range of ±0.5 D; the second group consists of patients for whom the calculated IOL values exceed the range of ±0.5 D but fall within the range of ±1.0 D when calculated using the second - main ANN model; the third group consists of patients whose calculated values exceed the range of ±1.0 D when calculated using the second - main ANN model. The second ANN model is the main one of the calculating ones, it takes into account linear, quadratic and cubic components, and its formula has the following form: , where x i are the values of the input variables of the ANN model: x 1 - refraction of the strong meridian, D; x 2 - refraction of the weak meridian, D; x 3 - optical eye length, mm; x 4 - anterior chamber depth, mm; x 5 - lens thickness, mm; x 6 - a-constant of the model of the implanted IOL; Y - output variable: value of IOL optical power, D; w i - coefficients of the ANN model. EFFECT: increasing the accuracy of preoperative IOL calculation. 2 cl, 3 dwg

Inventors

  • ARZAMASTSEV ALEKSANDR ANATOLEVICH
  • Fabrikantov Oleg Lvovich
  • Zenkova Natalya Aleksandrovna
  • Khudyakov Artem Aleksandrovich

Dates

Publication Date
20260506
Application Date
20250307

Claims (18)

  1. 1. A method for calculating the optical power of an intraocular lens (IOL) using the capabilities of artificial neural networks (ANN models) with preliminary classification of patient groups, characterized in that four ANN models are used, the first of which acts as a classifier and directs the patient, based on the data of his preoperative examination: refraction of the strong meridian, refraction of the weak meridian, optical length of the eye, depth of the anterior chamber, thickness of the lens, the a-constant of the model of the implanted IOL, to one of three calculation ANN models, each of which has its own architecture and calculates the optical power of the IOL only for the group of patients obtained during classification using the first ANN model, which is a fully connected unidirectional perceptron with four hidden layers and the following number of inputs and outputs of the neural network: the first layer - six inputs and 144 outputs; the second layer - 144 inputs and 72 outputs; the third layer - 72 inputs and 72 outputs; the fourth layer - 72 inputs and 3 outputs;
  2. The first group consists of patients for whom the calculated IOL values do not exceed the range of ±0.5 D; the second group consists of patients for whom the calculated IOL values exceed the range of ±0.5 D, but fall within the range of ±1.0 D when calculated using the second, main ANN model; the third group consists of patients for whom the calculated values exceed the range of ±1.0 D when calculated using the second, main ANN model; the second ANN model is the main one of the calculations, it takes into account the linear, quadratic and cubic components, and its formula is as follows:
  3. ,
  4. where x i are the values of the input variables of the ANN model:
  5. x 1 - refraction of the strong meridian, diopters;
  6. x 2 - refraction of the weak meridian, diopters;
  7. x 3 - optical length of the eye, mm;
  8. x 4 - anterior chamber depth, mm;
  9. x 5 - lens thickness, mm;
  10. x 6 - a-constant of the implanted IOL model;
  11. Y is the output variable: the value of the IOL optical power, diopters;
  12. w i - coefficients of the ANN model;
  13. The model is trained using the average relative error as the loss function; the stochastic gradient method, the coordinate descent method, and the Monte Carlo method are used as training methods;
  14. The third ANN model is a fully connected unidirectional perceptron with four hidden layers and the following number of inputs and outputs of the neural network: the first layer - six inputs and 240 outputs; the second layer - 240 inputs and 120 outputs; the third layer - 120 inputs and 60 outputs; the fourth layer - 60 inputs and 1 output;
  15. The fourth ANN model is a fully connected one-way perceptron with four hidden layers and the following number of inputs and outputs of the neural network: the first layer - six inputs and 240 outputs; the second layer - 240 inputs and 120 outputs; the third layer - 120 inputs and 12 outputs; the fourth layer - 12 inputs and 1 output;
  16. Next, the second main calculation ANN model is constructed and subjected to deep machine learning, and the trained ANN model is validated on new data; three different datasets are selected from the entire set of empirical data; the first includes only the rows for which the calculated values of the IOL optical power are within the range of ±0.5 D; the second includes only the rows for which the calculated values of the IOL optical power are not within the range of ±0.5 D, but are within the range of ±1.0 D; the third includes only the rows for which the calculated values of the IOL optical power are not within the range of ±1.0 D;
  17. then the third ANN model is trained on the second dataset, the fourth ANN model is trained on the third dataset, after which a dataset is compiled for training the first ANN model and the first ANN model - classifier is trained on the obtained dataset, after which it is possible to use this method for calculating the optical power of the IOL by entering the data of the preoperative examination of patients at the input of the first ANN model, which classifies them into the desired group, and then the IOL calculation is carried out using one of the three ANN models - the second, or one of the two auxiliary ones - the third or the fourth.
  18. 2. The method according to paragraph 1, characterized in that the ANN model is retrained by preparing a new training sample as new empirical data about patients of the ophthalmology clinic accumulates.

Description

The invention relates to the field of medicine, namely to ophthalmology, and can be used in calculating the optical power of an intraocular lens (IOL) using mathematical models obtained as a result of deep learning of artificial neural networks (ANN models). At the present stage, ophthalmic surgery involves achieving the maximum functional result, including the use of formulas for calculating the optical power of the IOL and obtaining a target refraction within ±0.5 D, since IOL manufacturers produce them with exactly this increment. One of the tools for optimizing the IOL calculation process is working on formulas for calculating the optical power of the IOL. There are a significant number of formulas for calculating the optical power of IOLs of the first to third generations, based both on the laws of geometric optics and on approximation dependencies for a significant number of patients: these are the formulas of Fedorov, SRK II, Haigis, Holladay, SRK-T and others (Balashevich L.I., Danilenko E.V. Results of using the formula of S.N. Fedorov for calculating the power of posterior chamber intraocular lenses // Ophthalmosurgery. - No. 1. - 2011; Sanders, D. R., Retzlaff, J. A. and Kraff, M. C. Comparison of the SRK-2 formula and other second-generation formulas. Journal of Cataract & Refractive Surgery. - 1988 (14), 136-141; Hoffer, K. J. The Hoffer Q formula: a comparison of theoretical and regression formulas. Journal of Cataract & Refractive Surgery. - 1993 - (19), 700-712). A common disadvantage of such formulas is a significant average relative error in calculating the IOL optical power (11-12%), a low percentage of the calculated value falling within the range of ±0.5 D (14-20%) (Arzamascev, A.A. Optimization of formulas for calculating IOL / A.A. Arzamascev, O.L. Fabrikantov, N.A. Zenkova, N.K. Belousov // Bulletin of Tambov University. Series: Natural and Technical Sciences. - 2016. - Vol. 21. - Issue 1. - pp. 208-212), and the fact that they work well for the "average patient", but are not adequate enough at the boundaries of the ranges of input variables. Their other disadvantages are the inability to take into account the non-stationarity of the object and adjustment when new empirical data are received, for example, in the case of their localization, as well as a small, clearly insufficient number of input factors taken into account. These circumstances give rise to a large number of local amendments to these formulas, their constant adaptation and correction. In recent years, the use of machine learning and the capabilities of neural networks to modernize and optimize the IOL calculation process has become increasingly relevant (Vinogradov A.R., Dzhashi B.G., Yuferov O.V., Balalin S.V., Tarapatina E.S. Modern possibilities for optimizing the calculation of the optical power of an intraocular lens using the capabilities of deep machine learning // Ophthalmosurgery. - No. 4S. - 2022. - p. 138). The most modern approach to calculating the optical power of IOLs is the use of multiparametric formulas of the fourth and fifth generations, which have the ability to take into account complex nonlinear relationships between the anatomical parameters of the eye and the required optical power of the IOL, based on the full or partial use of artificial intelligence systems. These are the Barrett Universal II, Hill-RBF, Kane, and Pearl DGS formulas (Stopyra, W., Cooke, D.L., Grzybowski, A. A. Review of intraocular lens power calculation formulas based on artificial intelligence // Journal of Clinical Medicine. - 2024, 13, 498. https://doi.org/10.3390/jcml3020498; Yamauchi T., Tabuchi T., Takase K., Masumoto H. Use of a machine learning method in predicting refraction after cataract surgery // Journal of Clinical Medicine. 2021, 10, 1103. doi: 10.3390/jcml0051103). These formulas are freely available as calculators. However, the lack of the ability to study their structure and coefficients makes their use inaccessible for individual calculations and excludes their integration into other decision support systems in ophthalmology. It is known that ANN models are used for data generalization and prediction of the optical power of modern IOLs, based on the expansion of the function of a large number of variables in a Taylor series, which have a significantly larger number of input variables and reflect to a much greater extent the regional specificity of patients than traditionally used formulas (Arzamastsev A.A., Fabrikantov O.L., Zenkova N.A., Belikov S.V. Application of machine learning technology to predict the optical power of intraocular lenses: generalization of diagnostic data // Digital Diagnostics. 2024. Vol. 5, No. 1. Pp. 53-63. DOI: https://doi.org/10.17816/DD623995). We have adopted this method as a prototype. However, its use does not allow us to obtain a level of average relative calculation error below 3.5%, which cannot be considered an optimal value for ensuring the possibility of obtaining the