TL;DR
This paper introduces a deep learning model that can accurately predict a patient's visual field (VF) from optical coherence tomography (OCT) scans, bypassing the need for traditional, often unreliable, perimetry tests. This breakthrough could revolutionize glaucoma diagnosis and monitoring, especially for patients who struggle with current VF testing methods.
What Problem Does This Paper Address?
Glaucoma is a leading cause of irreversible blindness, characterized by progressive damage to the optic nerve and retinal ganglion cells. Monitoring the functional loss of vision, or visual field (VF), is crucial for managing glaucoma. The current gold standard for this is Standard Automated Perimetry (SAP), typically performed with devices like the Humphrey Field Analyzer (HFA). However, SAP suffers from significant intra-subject variability and can be time-consuming and difficult for some patients to complete reliably, especially those with advanced disease or cognitive impairments.
Optical Coherence Tomography (OCT) provides detailed structural information about the retina, including the thickness of various layers like the Retinal Nerve Fiber Layer (RNFL). While there's a known relationship between these structural measures and VF loss (the "structure-function" relationship), directly translating OCT scans into accurate VF predictions has been challenging. Previous attempts often relied on manual segmentation of retinal layers or made simplifying assumptions about this relationship, limiting their accuracy. This paper aims to bridge this gap by developing a deep learning model that can directly predict VF from unsegmented OCT scans, offering a more robust and potentially faster alternative to SAP.
Methods
The researchers developed and evaluated eight deep learning (DL) regression models. These models were trained using different retinal imaging modalities: circumpapillary OCT scans at three different diameters (3.5mm, 4.1mm, and 4.7mm) and Scanning Laser Ophthalmoscopy (SLO) en face images. The models were designed to predict two key metrics: the Mean Deviation (MD), an overall measure of VF loss, and 52 individual VF threshold values, representing specific points on the visual field map.
A key methodological innovation is the use of unsegmented OCT scans. This means the DL model directly processes the raw voxel data from the OCT, avoiding the need for laborious and potentially error-prone manual or automated segmentation of retinal layers. The researchers employed an Xception architecture, pre-trained on ImageNet, which was then adapted for regression tasks. They explored both individual models and ensemble (averaged) predictions from multiple models to improve performance. Training involved optimizing for Mean Squared Error (MSE) loss using the Adam optimizer, with data augmentation techniques like horizontal flipping, elastic deformation, and cutout applied to enhance model robustness.
Datasets
The study utilized a retrospective dataset of 1378 matched OCT-VF pairs from 496 patients (863 eyes) who had undergone a complete glaucoma examination, including a reliable Humphrey Field Analyzer (HFA) 24-2 SITA Standard VF exam and a SPECTRALIS OCT scan.
The data was split into training, validation, and testing sets. For the validation and test sets, standard HFA reliability indices (false positives, false negatives, and fixation losses) were used to filter out unreliable VF data. The training set, however, included unreliable data as a form of data augmentation. No new benchmark datasets were introduced; instead, the study leveraged existing clinical data to validate its approach against established VF measurements.
Key Insights and Ideas
- Unsegmented OCT is Sufficient: The study demonstrates that deep learning models can accurately predict visual field parameters directly from raw, unsegmented OCT data. This bypasses the complex and error-prone step of retinal layer segmentation, simplifying the pipeline and potentially improving robustness.
- OCT Outperforms SLO for VF Prediction: Circumpapillary OCT scans consistently outperformed SLO en face images in predicting both MD and individual VF thresholds. This suggests that the volumetric and cross-sectional information provided by OCT is more informative for structure-function mapping than the surface-based SLO images.
- Ensemble Models Boost Performance: Averaging the predictions from multiple DL models (specifically, the circumpapillary OCT models) led to improved accuracy for MD estimation, highlighting the benefit of model ensembling.
- Larger OCT Rings Show Slight Advantage: For individual circumpapillary OCT scans, the 4.7mm diameter ring generally showed slightly better performance than smaller rings, possibly because it captures a broader area of the optic nerve head and surrounding retinal structures.
- Consistent Prediction Across VF Severity: The DL models demonstrated consistent performance across different levels of VF severity, with error ranges comparable to or better than previous non-DL studies. This is crucial for reliable monitoring across the disease spectrum.
Key Results
- MD Estimation: The best ensemble model, averaging predictions from the three circumpapillary OCT scans, achieved a Pearson correlation coefficient (r) of 0.79 on the independent test set for Mean Deviation (MD) estimation. This represents a significant improvement over models trained solely on SLO images (r=0.68).
- Pointwise VF Threshold Estimation: The best model for predicting individual VF threshold values (using the 4.7mm OCT scan) achieved a Pearson r of 0.77 on the validation set and 0.79 on the test set. The ensemble of three circumpapillary OCT scans further improved this to 0.78 on the validation set and 0.79 on the test set.
- Error Reduction: The DL models significantly reduced the baseline Mean Absolute Error (MAE) for MD estimation by 53% and for VF threshold estimation by 38%.
- Test-Retest Comparison: The study found that 88% of the predicted VF sensitivity error ranges fell within the 90% confidence intervals of empirically established test-retest variability, indicating a high degree of consistency and reliability in the model's predictions.
Limitations and Future Work
- Single-Center, Single-Device Data: The models were trained and validated on data from a single hospital using a single type of OCT device and VF exam. The authors acknowledge the need to further validate these models on external datasets from different OCT devices and patient populations to assess generalizability.
- Lack of Explainability: While the data-driven approach avoids segmentation errors, it comes at the cost of model transparency. The authors plan to investigate which retinal layers or features are most informative for the DL-predicted VF maps in future studies.
- Generalizability: The generalizability of the models to different OCT devices and VF testing protocols is not yet fully established.
- Future Work: The authors envision that this technology could become a valuable tool for visual function estimation in patients who cannot perform reliable VF exams, potentially offering a faster and more consistent alternative to current methods. They also plan to explore the informativeness of different retinal layers for DL-predicted VF maps.
Summary updated