Glaucoma detection beyond the optic disc: The importance of the peripapillary region using explainable deep learning

Ruben Hemelings, Bart Elen, João Barbosa Breda, Matthew B. Blaschko, Patrick De Boever, Ingeborg StalmansarXiv 2021 · DOI · OpenReview

TL;DR

This paper introduces a novel explainable deep learning approach to detect glaucoma and estimate the vertical cup-to-disc ratio (VCDR) from fundus images. By systematically cropping images to isolate or exclude the optic nerve head (ONH) and its surrounding peripapillary region, the study demonstrates that deep learning models can effectively identify glaucoma even when the ONH is completely removed, highlighting the importance of extra-ONH information.

What Problem Does This Paper Address?

Glaucoma is a leading cause of irreversible blindness, and early detection is crucial. While deep learning models have shown impressive accuracy in detecting glaucoma from fundus images, they often lack transparency, making it difficult to understand why they make a particular diagnosis. This "black box" nature hinders clinical trust and adoption. Furthermore, current models primarily focus on the optic nerve head (ONH), which is where characteristic glaucoma damage is traditionally assessed. However, it's unclear how much information lies in the peripapillary region (surrounding the ONH) and whether deep learning can leverage this information, especially in cases where ONH assessment might be challenging due to anatomical variations. This paper aims to bridge this gap by investigating the role of peripapillary regions in glaucoma detection and VCDR estimation using explainable AI techniques.

Methods

The core of this research lies in a systematic "cropping policy" applied to fundus images. Two main strategies were employed:

  • ONH Crop Policy: Images were cropped to progressively remove the optic nerve head (ONH) and surrounding areas, starting from 10% of the image size up to 60%. This allowed the researchers to assess performance when ONH information was gradually removed.
  • Periphery Crop Policy: This policy involved cropping from the periphery inwards, effectively removing the outer parts of the image while keeping the ONH intact. This was used to understand the contribution of the peripapillary region when the ONH is present.

Crucially, the researchers explored two training paradigms: retraining the models on the cropped images and no retraining (occlusion testing). Retraining allowed the model to adapt to the new image composition, while no retraining tested how well a model trained on full images performed when parts were masked.

For explainability, saliency maps were generated. These maps highlight the image regions that the deep learning model found most important for its predictions, providing visual insights into the decision-making process.

Datasets

The study utilized two primary datasets:

  • UZL Dataset: A retrospective dataset from the University Hospitals Leuven (UZL) in Belgium, comprising 37,627 stereoscopic color fundus images from 9,965 patients. This dataset was used for both VCDR regression and glaucoma classification. For VCDR regression, 23,930 images with valid VCDR labels were used. For glaucoma classification, 13,551 images were selected.
  • REFUGE Dataset: A publicly available dataset from the Retinal Fundus Glaucoma Challenge (REFUGE), consisting of 1,200 fundus images. This dataset was used to validate the generalization capabilities of the models trained on the UZL data.

Key Insights and Ideas

  • Glaucoma detection is possible without the ONH: Perhaps the most striking finding is that deep learning models can still achieve significant performance in glaucoma detection even when the ONH is completely removed from the image. This challenges the traditional clinical focus solely on the ONH.
  • Peripapillary region is crucial: The study strongly suggests that the peripapillary region, surrounding the ONH, contains vital information for glaucoma diagnosis and VCDR estimation.
  • Explainability reveals new patterns: Saliency maps revealed that important features for glaucoma detection and VCDR estimation are often located in the superotemporal and inferotemporal sectors outside the ONH, areas known to have the thickest retinal nerve fiber layer (RNFL).
  • Retraining enhances robustness: Models that were retrained on cropped images (even with significant ONH removal) performed substantially better than those tested without retraining, indicating the model's ability to adapt and learn from new image compositions.
  • VCDR estimation is more ONH-dependent than glaucoma detection: While both tasks benefit from the ONH, VCDR estimation showed a more pronounced performance drop when the ONH was removed compared to glaucoma detection.

Key Results

  • Glaucoma Detection:
    • Models trained on original images achieved an AUC of 0.94 [95% CI: 0.92-0.96] on the UZL test set.
  • Even with 60% of the image cropped to remove the ONH, models still achieved an AUC of 0.88 [95% CI: 0.85-0.90] when retrained.
  • The REFUGE dataset validation showed similar generalization capabilities, with AUCs around 0.87 for original images and 0.80 at 60% ONH crop.
  • VCDR Estimation:
    • Models trained on original images achieved an R² of 77% [95% CI: 0.77-0.79] and a Pearson correlation of 0.88.
  • With 60% ONH crop and retraining, the model still explained 37% [95% CI: 0.35-0.40] of the variance, with a moderate correlation of 0.62 [95% CI: 0.60-0.63].
  • The mean absolute error (MAE) for VCDR prediction was significantly reduced compared to a baseline prediction.

Limitations and Future Work

  • Fixed Crop Size: The study used fixed-size masks for cropping, which might introduce variations due to differences in ONH size across the population. Future work could define crop radii relative to the optic disc size.
  • Myopia Influence: The influence of myopic changes (e.g., peripapillary atrophy, tilted discs) was not explicitly assessed due to data limitations. Future studies could incorporate this information.
  • Disease Stage: The role of disease stage was not analyzed. Future experiments could investigate if deep learning explainability aligns with the known progression of glaucoma damage (infero-temporal and supero-temporal).
  • Glaucoma Suspects: The study focused on confirmed glaucoma cases to ensure clear explainability. Extending the approach to glaucoma suspects would be valuable.

The authors suggest future research could explore the impact of myopia, analyze disease stages, and investigate glaucoma suspects to further refine the clinical applicability of their explainable deep learning methods.

Summary updated

Abstract

Although unprecedented sensitivity and specificity values are reported, recent glaucoma detection deep learning models lack in decision transparency. Here, we propose a methodology that advances explainable deep learning in the field of glaucoma detection and vertical cup-disc ratio (VCDR), an important risk factor. We trained and evaluated deep learning models using fundus images that underwent a certain cropping policy. We defined the crop radius as a percentage of image size, centered on the optic nerve head (ONH), with an equidistant spaced range from 10%-60% (ONH crop policy). The inverse of the cropping mask was also applied (periphery crop policy). Trained models using original images resulted in an area under the curve (AUC) of 0.94 [95% CI: 0.92-0.96] for glaucoma detection, and a coefficient of determination (R^2) equal to 77% [95% CI: 0.77-0.79] for VCDR estimation. Models that were trained on images with absence of the ONH are still able to obtain significant performance (0.88 [95% CI: 0.85-0.90] AUC for glaucoma detection and 37% [95% CI: 0.35-0.40] R^2 score for VCDR estimation in the most extreme setup of 60% ONH crop). Our findings provide the first irrefutable evidence that deep learning can detect glaucoma from fundus image regions outside the ONH.
BibTeX
@misc{hemelings2021glaucoma,
  author = {Ruben Hemelings and Bart Elen and João Barbosa Breda and Matthew B. Blaschko and Patrick De Boever and Ingeborg Stalmans},
  title = {{Glaucoma detection beyond the optic disc: The importance of the peripapillary region using explainable deep learning}},
  year = {2021},
  eprint = {2103.11895},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2103.11895},
}
Figure