Artificial Intelligence Based Grading Of Nuclear And Posterior Subcapsular Cataract Using Oct Images

To develop and validate an artificial intelligence (AI) algorithm for automatic nuclear and posterior subcapsular cataract (PSC) grading using optical coherence tomography (OCT) scans of the human lens.

Dardenne Eye Hospital, Bonn, Germany

We developed an AI algorithm using convolutional neural networks based on 1760 OCT images of cataractous lenses from 880 patients (mean age: 72 years; range: 26-92 years). The lenses were expert-labelled using the LOCS III grading system. Nuclear cataract grades were classified as mild (≤3.0), moderate (>3.0 and ≤4.0), and severe (>4.0). Posterior subcapsular cataract (PSC) was dichotomized as "moderate" for grades <4 and "severe" for ≥4. The images were randomly divided into training (70%) and test (30%) sets.

The developed algorithm achieved an accuracy of 61.3% with "near miss" false predictions of 32.4% for nuclear cataract and an accuracy of 83.87% for PSC, respectively. The class activation maps corresponded to the areas of highest interest when assessing lens opacity by slitlamp microscopy.

Our deep learning algorithm enables automated grading of nuclear and posterior subcapsular cataracts on OCT images of the lens without the need for preprocessing. Our tool may facilitate machine-based clinical analysis of cataractous lenses.