Location
Philadelphia, PA
Start Date
1-5-2024 1:00 PM
End Date
1-5-2024 4:00 PM
Description
Precise intraocular lens (IOL) selection plays a critical role in optimizing surgical outcomes for patients undergoing IOL surgery. Traditionally, only the anterior corneal astigmatism (ACA) was considered during IOL selection. However, the posterior corneal astigmatism (PCA), also significantly impacts vision, though it is more challenging to measure directly. Advancements in deep learning offer a promising solution. By analyzing existing ACA measurements, deep learning algorithms can predict total corneal astigmatism (TCA) with high accuracy, effectively inferring PCA and enabling more precise IOL selection.
We employed a multivariate linear regression neural network (NN) and evaluated its performance through R-squared and root mean square error (RMSE) which indicates the error between our actual and predicted values. To predict the TCA magnitude in diopters, the NN was trained on three key patient characteristics: age, ACA magnitude, and ACA axis. Predicting the TCA axis involved a two-step process. First, vectorized components of TCA magnitude were used to obtain a predicted magnitude and axis. We compared the effectiveness of three different input processing methods: the raw untransformed dataset, the Cravy method, and the Humphrey method. The Cravy and Humphrey methods allowed us to vectorize ACA into X and Y components.
Our model achieved a root mean square error (RMSE) of 0.1013 compared to the formula's 0.2258 for TCA magnitude. Additionally, our model achieved a higher R-squared value of 0.972, indicating stronger correlation between predicted and actual values compared to the formula's 0.8624. Our deep learning approach demonstrated significantly fewer deviations from the actual values as the error magnitude increased. This implies our model can consistently predict the axis with greater accuracy across a wider range of values.
By applying deep learning techniques, we significantly improved the accuracy of TCA prediction, leading to better selection of toric IOLs and more precise PCA calculations.
Embargo Period
6-13-2024
Included in
Improving IOL Calculators’ Total Corneal Astigmatism Predictions with Deep Learning
Philadelphia, PA
Precise intraocular lens (IOL) selection plays a critical role in optimizing surgical outcomes for patients undergoing IOL surgery. Traditionally, only the anterior corneal astigmatism (ACA) was considered during IOL selection. However, the posterior corneal astigmatism (PCA), also significantly impacts vision, though it is more challenging to measure directly. Advancements in deep learning offer a promising solution. By analyzing existing ACA measurements, deep learning algorithms can predict total corneal astigmatism (TCA) with high accuracy, effectively inferring PCA and enabling more precise IOL selection.
We employed a multivariate linear regression neural network (NN) and evaluated its performance through R-squared and root mean square error (RMSE) which indicates the error between our actual and predicted values. To predict the TCA magnitude in diopters, the NN was trained on three key patient characteristics: age, ACA magnitude, and ACA axis. Predicting the TCA axis involved a two-step process. First, vectorized components of TCA magnitude were used to obtain a predicted magnitude and axis. We compared the effectiveness of three different input processing methods: the raw untransformed dataset, the Cravy method, and the Humphrey method. The Cravy and Humphrey methods allowed us to vectorize ACA into X and Y components.
Our model achieved a root mean square error (RMSE) of 0.1013 compared to the formula's 0.2258 for TCA magnitude. Additionally, our model achieved a higher R-squared value of 0.972, indicating stronger correlation between predicted and actual values compared to the formula's 0.8624. Our deep learning approach demonstrated significantly fewer deviations from the actual values as the error magnitude increased. This implies our model can consistently predict the axis with greater accuracy across a wider range of values.
By applying deep learning techniques, we significantly improved the accuracy of TCA prediction, leading to better selection of toric IOLs and more precise PCA calculations.
Comments
Winner of 2024 PCOM David Miller, DO ’60 Endowed Memorial Research Day Best in Show Award.