Abstract
Purpose :
Visual field (VF) interpretation in glaucoma management relies traditionally on subjective assessment which lacks consistency and efficiency. Archetypal Analysis (AA), an unsupervised learning approach, has demonstrated effectiveness in quantifying distinct VF patterns in glaucomatous loss [1]. We explored if using AA as a feature extraction layer could improve deep learning classification of glaucomatous VF defects.
Methods :
We applied AA to total deviation (TD) values of 3814 24-2 VFs from 1692 eyes (mean age 63 years, SD=16; average mean deviation -6.5dB, SD=7.5), with augmentation via vertical flipping. AA weights were extracted from a subset of 5612 VFs (90% training, 10% test), including healthy VFs and those marked with one or more common glaucomatous defects (N = 10) by two glaucoma specialists, with an adjudication process for resolving discrepancies. The AA weights were then used as input to a multilayer perceptron (MLP) to classify the different glaucomatous defects in a multilabel format. We compared our AA+MLP model to similar MLP architectures fit to raw TD values (raw+MLP) and weights derived from an AA model using 13231 VFs developed by Elze et al [1] (AAElze+MLP). The MLP models were evaluated using the macro-area under the curve (AUC) metric, applied to the test set.
Results :
AA determined 16 distinct archetypes (ATs). Fig 1 shows the top three ATs for two VFs decomposed by our AA model (black) and the AAElze model (red), revealing some similarities in AT patterns despite the models being trained on distinct datasets. We observed superior macro-AUC scores for the AA-based MLP models, 0.906 for AA+MLP and 0.914 for AAElze+MLP compared with 0.889 for the raw+MLP (Fig 2). Notably, raw+MLP exhibited increased variability in its performance across different classes of VF defects, particularly the central and paracentral defects, suggesting the enhanced stability of AA-based MLP models in the classification of glaucomatous VF defects.
Conclusions :
AA or other unsupervised approaches may help refine the classification of visual field defects, thereby supporting more consistent and efficient glaucoma management. With a larger dataset, these approaches may improve the performance and reliability of glaucomatous defect classification or optimize the model-building workflow by streamlining the selection of data for annotation from extensive, unlabeled datasets.
[1] Elze et al.J Royal Soc In 2015; 12(103)
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.