Machine and deep learning predictions of visual fields from spectral-domain optical coherence tomography retinal nerve fiber layer thickness maps in glaucoma vision loss

Abstract

Purpose : This study develops and evaluates traditional machine learning (ML) and deep learning methods to predict functional visual fields (VF) from spectral domain optical coherence tomography (SD-OCT) retinal nerve fiber layer thickness (RNFLT) maps for various stages of glaucoma. Firstly, the traditional circumpapillary circle scan of the RNFLT map as an input into the traditional ML model is compared with using the entire thickness map using a Convolutional Neural Network (CNN) model. Secondly, the inclusion of the Optic Nerve Head (ONH) in the thickness map is investigated to understand if this improves the CNN predictions.

Methods : The dataset comprised OCT RNFLT maps and VF scans of 56,455 eyes from the Mass. Eye and Ear clinic. Normalized RNFLT maps with centered ONHs were used as optimized inputs. CNN and traditional ML models were independently trained to predict mean deviation (MD) from RNFLT maps with regression outputs. VGG-16 was used as the base CNN architecture and the encoder output was followed by a global average pooling and 2 levels of dense layers. ‘ReLu’ action was used for the first level and ‘Linear’ activation for the second. Random Forest Regressor (RFR) was used for the traditional ML model. The ML models were trained using the circumpapillary circle scan of the RNFLT map. During CNN training 2 different inputs were considered: (1) RNFLT maps and (2) RNFLT maps with the ONH omitted (see Fig. 1). The rationale for removing the ONH is that, physiologically, there is no RNFL present in this region. However, the information at the ONH edge may still be useful in improving the model training so that it can deduce the ONH redundancy in terms of RNFL.

Results : Fig. 2 presents the MAE between the actual and predicted MD values to evaluate the model performance. Overall, the CNN model using the ONH omitted outperforms the other approaches, whereas the use of the full RNFLT map into the CNN demonstrates the lowest overall performance.

Conclusions : The deep learning model predictions are most accurate when the RNFLT maps are filtered to omit the ONH region. Unexpectedly, the RFR predictions do not demonstrate the highest overall MAE, despite utilizing comparatively limited information on the RNFLT as input. In future, the features identified by the CNNs with respect to their clinical relevance will be investigated.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.