A method for quantifying sectoral optic disc pallor in fundus photographs and its association with peripapillary RNFL thickness

Emanuele Trucco, Frsa, Fiapr, Fbmva, Faaia, Samuel Gibbon, Graciela Muniz-Terrera, Fabian S. L. Yii, Charlene Hamid, Simon Cox, Ian J. C. MacCormick, Andrew J. Tatham, Craig Ritchie, Emanuele Trucco, Baljean Dhillon, Thomas J. MacGillivrayarXiv 2023 · DOI · arXiv

This research presents a significant advancement in making objective ophthalmic screening more accessible. By combining deep learning with traditional colorimetric analysis, the authors have developed a tool that transforms standard fundus photographs into quantitative maps of optic nerve health, providing a low-cost alternative to specialized imaging.

The Challenge: Subjectivity in Optic Atrophy

Optic disc pallor—a loss of the healthy reddish hue of the optic nerve head—is a primary clinical sign of optic atrophy, indicating irreversible damage to retinal ganglion cell axons. While clinicians routinely look for this during examinations, the assessment is notoriously subjective. Two doctors may disagree on the degree of pallor, and tracking subtle changes over time is nearly impossible with the naked eye.

Comparison between fundus photography (A, E) and the gold standard OCT (B, C, D) for assessing retinal nerve fiber layer (RNFL) health.

The current gold standard for measuring this damage is Optical Coherence Tomography (OCT), which measures the thickness of the Retinal Nerve Fiber Layer (RNFL). However, OCT machines are expensive, require specialized training, and are less prevalent in primary care or low-resource settings. This study addresses the need for an automated, objective, and low-cost way to quantify nerve health using standard fundus cameras.

Methods: An Automated Pipeline

The researchers developed a multi-stage pipeline using DeepLabv3+, a state-of-the-art deep learning architecture for image segmentation. The process involves four key automated tasks: localization of the optic disc, precise segmentation of the disc margin, fovea localization to establish a "fovea-disc axis," and vessel segmentation to exclude blood vessels from color analysis.

Overview of the automated pipeline, from dataset selection to the final pallor software.

The "Secret Sauce"

Unlike previous methods, this software rotates the image based on the fovea-disc axis. This allows the software to divide the disc into "sectors" that correspond exactly to the anatomical zones used in OCT reports, enabling a direct "apples-to-apples" comparison between a 2D photograph and a 3D OCT scan. Pallor is calculated as the ratio of green to red pixel intensities within the neuro-retinal rim, normalized against a control region of the background retina.

Datasets and Robustness

The study utilized seven diverse datasets (including PREVENT, LBC 1936, ORIGA, and FIVES) to ensure the AI could handle real-world variability. The software was intentionally tested on images with varying resolutions, file formats, and camera manufacturers (Canon, Topcon, Zeiss). It maintained high performance across the board, suggesting it is ready for clinical data from diverse environments.

Key Results

  • Efficiency: The entire pipeline processes a single image in less than 3 seconds.
  • Correlation with Gold Standard: Automated pallor measurements showed a statistically significant association with global pRNFL thickness ($\beta = -9.81, p < 0.05$). The strongest correlation was found in the temporal-inferior zone, a critical area for early disease detection.
  • Clinical Accuracy: When tested on the RFMiD dataset, the software successfully distinguished discs diagnosed as "pale" by ophthalmologists from healthy controls with high statistical significance ($p < 10^{-11}$).
  • Expert-Level Consistency: The AI's disc segmentation achieved a mean Intersection over Union (mIoU) of 0.959, rivaling the consistency of human experts.

Limitations and Future Work

The authors identify two primary hurdles for future refinement. First, media opacities like cataracts can change the color balance of a photograph, potentially creating "pseudo-pallor." Future versions will need to incorporate lens status as a covariate. Second, variations in individual skin and eye pigmentation can affect the results. Developing a "normative database" similar to those used in OCT would help account for this natural human variation.

Overall, this work provides a robust framework for bringing high-level diagnostic insights to standard fundus photography, potentially allowing for earlier detection of glaucoma and neurodegenerative diseases in primary care settings.

Summary updated

BibTeX
@misc{faaia2023method,
  author = {Emanuele Trucco, FRSA, FIAPR, FBMVA, FAAIA and Samuel Gibbon and Graciela Muniz-Terrera and Fabian S. L. Yii and Charlene Hamid and Simon Cox and Ian J. C. MacCormick and Andrew J. Tatham and Craig Ritchie and Emanuele Trucco and Baljean Dhillon and Thomas J. MacGillivray},
  title = {{A method for quantifying sectoral optic disc pallor in fundus photographs and its association with peripapillary RNFL thickness}},
  year = {2023},
  eprint = {2311.07213},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2311.07213},
}
Figure