Artificial Intelligence in Cornea, Refractive, and Cataract Surgery

Aazim A. Siddiqui; John G. Ladas; Jimmy K. Lee


Curr Opin Ophthalmol. 2020;31(4):253-260. 

In This Article

Artificial Intelligence and Cataract Surgery

As in keratoconus, artificial intelligence can be not only used for diagnostic purposes but also to guide surgical management. According to the WHO,[7] it is estimated that approximately 32 million cataract surgeries will have been performed on a yearly basis as of the year 2020.[7] It is the most commonly performed surgery in the United States and in the field of ophthalmology. Although the goal of cataract surgery is to improve a patient's overall visual function, appropriate consideration of IOL selection allows it to become the ultimate refractive surgery. This is achieved by meeting the desired refractive outcome. The aim is to maximize accuracy and minimize postoperative refractive error in order to achieve emmetropia.

The resulting two-lens system of the cornea and the intraocular lens (IOL), and the axial length of the eye, largely determines a given patient's postoperative refraction. Modern intraocular lens formulas take these factors from a given eye into account in order to calculate a lens power, which corresponds to a targeted refraction. For a given eye, the process of choosing an appropriate IOL is mostly guided by an individual surgeon's formula preferences and experience.

With increasing demands, a generally accepted goal for a postoperative refractive outcome is to be within 0.50 D of emmetropia or slight myopia. However, in reality, this goal is only achieved approximately 70–80% of the time with any single unoptimized formula.[8] This leaves approximately one out of four patients with a refractive outcome that is greater than 0.50 D off of targeted refraction. This refractive variance may then lead to a need for significant spectacle correction or further surgery to address remaining refractive error.

Since the early days of the first-generation and second-generation IOL formulas, the field of lens calculations has evolved greatly. There now exist several methodologies of optimizing and improving the processes of intraocular lens calculations. However, there continues to be a lack of a single, perfect IOL calculation solution that has been demonstrated to be the most accurate for all types of eyes. Therefore, there still remains a gap in achieving the optimal postoperative refraction for a significant portion of patients.[9]

As a result of this paucity, most surgeons rely on results from multiple IOL formulas to choose the most appropriate lens for their patients, thus requiring a significant dedication of clinical workflow and time. The desire to simplify this process and arrive at a more accurate IOL calculation framework has led to the development of the next generation of IOL formulas and integration of these formulas with artificial intelligence.

Introduction of theoretical formulas, such as the Holladay, SRK/T, Hoffer Q, and Haigis provided an improved level of accuracy for many years over the early linear regression formulas, such as the SRK I and SRK II. Recently, there has been an influx of newer generation of formulas: Barrett, Olsen, Holladay 2, and the Hill-RBF formulas. Further, many of these formulas have been optimized in various ways to help achieve greater accuracy. Despite these improvements and advancements, each formula continues to have limitations under certain situations.[9]

A significant step towards the integration of artificial intelligence in IOL calculations took place in 2015 with the introduction of a concept of an IOL 'super formula'.[10] Although previous generations of IOL formulas were thought of as two-dimensional algebraic equations, this novel methodology depicted formulas in three dimensions. This allowed for a 3-D analysis framework to observe areas of similarities and disparities between IOL formulas. Next, best portions of each of the modern IOL formulas were chosen and an IOL 'super surface' was developed based on an amalgam of these formulas (Figure 1 a-c). From this super surface, the 'super formula' was derived.

Figure 1.

Super surface and super formula. (a) The initial super surface was constructed based on the Hoffer Q, Holladay I, Holladay I with the Koch adjustment, and Haigis formulas. (b) Sections of each formula joined together. (c) Formulas are amalgamated into one super surface.

The super formula may be used to calculate eyes of average axial length and keratometry values. Further, given that it is an amalgam of various other formulas, it also performs well for eyes with short or long axial lengths, steep or flat corneas, and shallow or deep anterior chambers. Figure 2 a demonstrates an example calculation of an 'average' eye with axial length and keratometry values within a typical range – the super formula localizes to the correct region of the super surface and provides an IOL power value. Next, an eye with a shorter axial length is demonstrated in Figure 2 b. It is known that an eye with shorter axial length is difficult to calculate given that the smallest of shifts in the effective lens position may result in dramatic changes in the refractive outcome – this can be observed graphically in Figure 1 c as the surface is steeper in areas of shorter axial length. These examples highlight the potential applications and versatility of the super formula.

Figure 2.

Super formula examples. (a) Sample calculation of an eye with axial length and keratometry values within a typical range. (b) Sample calculation of an eye with shorter axial length. Both images taken from online calculator.

The concept of three-dimensionality that is innate within the super formula and its depicted super surface serves as a graphical method of comparing one or more formulas. This allows for further evaluation of the graphical areas of clinical agreements and disagreements between multiple formulas. The areas of clinical disagreements between formulas may be examined and refined further. The use of this formula may take the burden away from the surgeon of having to select from multiple formulas for any given eye. The super formula may serve as a single solution to calculate eyes with both typical and atypical values of axial length, corneal power, and anterior chamber depth. Finally, it provides a blueprint that allows for adjustment and improvement of the formula.

The formula as published served as a dependable backbone to what is now an advanced version of this formula. In its original state, it used axial length, corneal power, anterior chamber depth, lens constant, and target refraction values as input parameters. With the help of deep learning techniques and artificial intelligence, this formula has evolved to an improved level of accuracy. This optimization process applies a unique adjustment to an eye based on its axial length, corneal power, and anterior chamber depth.

The latest iteration of the formula is optimized using actual postoperative outcomes taken from several high-volume cataract surgeons. Results from early studies based on this novel, hybrid approach using artificial intelligence have been promising. A recent small-scale study using an artificial intelligence-hybrid formula resulted in a significant improvement in a single surgeon's outcomes from 76% (using standard super formula) to 80% (using super formula with an artificial intelligence algorithm) of eyes reaching within 0.50 diopters of predicted refraction.[11] Further larger scale studies are ongoing to continue evaluation of this methodology.

Using this 'big data' methodology may further help achieve a higher level of accuracy in eyes with atypical parameters, such as those with a flat or steep cornea, long or short axial length, deep or shallow anterior chamber, or other parametric variations. Early iterations of this methodology has helped to significantly minimize the 'delta' between the predictions made by the formula and actual postoperative refractive outcomes.