Evaluation of the corneal topography based on deep learning

Shuai Xu, Xiaoyan Yang, Shuxian Zhang, Xuan Zheng, Zheng Fang, Yin Liu, Hanyu Zhang, Lihua Li, Qing YeFrontiers in Medicine 2024 · DOI

TL;DR

Researchers have developed a deep learning framework that automates the evaluation of corneal topography maps for patients using Orthokeratology (OK) lenses. By utilizing U-Net architectures to segment key medical zones, the system achieves over 98% accuracy in classifying topography types, significantly reducing the manual workload for optometrists while improving the precision of lens fitting.

What Problem Does This Paper Address?

Myopia (nearsightedness) is a global health crisis, with projections suggesting that 84% of children in China will be myopic by 2050. Orthokeratology (OK) lenses—specialized hard contacts worn overnight to reshape the cornea—are one of the most effective interventions for slowing myopia progression.

However, "fitting" these lenses is a complex, iterative process. Optometrists must analyze corneal topography (a color-coded map of the eye's surface) to see how well the lens is working. Currently, this analysis is largely manual, subjective, and time-consuming. Optometrists look for "bull’s-eye" or "smiley-face" patterns to determine if the lens is centered or if the "treatment zone" is correctly positioned over the pupil. This paper aims to replace this subjective "eyeballing" with an objective, automated deep learning pipeline.

Methods

The researchers developed a multi-stage approach combining deep learning with traditional computer vision:

  1. Semantic Segmentation: They compared three iterations of the U-Net architecture (U-Net, U-Net++, and U-Net3+) to identify two critical regions on topography maps: the pupil and the treatment zone.
  2. Feature Extraction: Using the OpenCV library, the system calculates specific medical indicators from the segmented masks:
  3. Decentration: The Euclidean distance between the center of the pupil and the center of the treatment zone.
  4. Effective Defocusing Contact Range (EDCR): The proportion of the pupil boundary that intersects with the treatment zone boundary.
  5. Classification: Based on these calculated metrics, the system automatically categorizes the topography into one of four classes (Class I to IV), ranging from "ideal" fits to those requiring immediate lens adjustment.

Datasets

The study utilized a retrospective dataset from the Tianjin Eye Hospital Optometry Center:

  • Total Scope: Clinical data from 1,302 myopic subjects (2,604 eyes) collected between 2013 and 2021.
  • Analysis Subset: 500 eyes were used for the primary statistical and AI analysis.
  • Training/Validation: 809 high-quality corneal topographies were manually labeled by experts to create the ground truth. The data was split into 80% for training and 20% for testing, with 10-fold cross-validation to ensure robustness.
  • Hardware: All images were generated using a Tomey corneal topographer.

Key Insights and Ideas

  • U-Net3+ Wins on Complexity: While the standard U-Net is faster and easier to deploy, U-Net3+ provided the best segmentation performance. The authors attribute this to its "full-scale skip links," which better preserve semantic information from the original image across different scales.
  • Objective vs. Subjective: The study highlights that human optometrists are prone to subjectivity and fatigue. By calculating the EDCR (a metric representing how much "defocus" signal the eye actually receives), the AI provides a more direct link to how well the myopia is being controlled than a simple visual inspection.
  • Speed Matters: The deep learning model takes only 0.06 seconds to segment the zones, and the entire pipeline runs in about 2.2 seconds. This is a massive improvement over the "half a minute or more" required for a manual expert review.

Key Results

  • Segmentation Accuracy: For the treatment zone, the U-Net3+ model achieved a Precision of 0.9587 and a Recall of 0.9378. For the pupil, it reached a Precision of 0.9771.
  • Classification Reliability: When compared against the "gold standard" labels provided by professional optometrists, the automated system demonstrated an accuracy of more than 98%.
  • Clinical Correlation: The study confirmed that Class I and II topographies (low decentration, high EDCR) were associated with slower axial elongation (eye growth) in younger children, validating the clinical importance of the AI's classification.

Limitations and Future Work

  • Brand Specificity: The model was trained exclusively on images from Tomey topographers. Because different manufacturers use different color scales and layouts, the model would likely need to be retrained or "fine-tuned" to work with other brands like Pentacam or Medmont.
  • Sample Size: While 1,302 subjects is a solid start, the authors note that a larger, more diverse sample size could further optimize the model and prevent overfitting.
  • Prospective Validation: As a retrospective study, the next logical step is a prospective clinical trial where the AI assists optometrists in real-time during the fitting process to see if it improves long-term myopia outcomes.

Summary updated

Abstract

The current study provided an effective and accurate deep learning algorithm to evaluate the type of corneal topography. The deep learning algorithm played an auxiliary role in the OK lens fitting, which could help optometrists select the parameters of OK lenses effectively.
BibTeX
@inproceedings{xu2024evaluation,
  author = {Shuai Xu and Xiaoyan Yang and Shuxian Zhang and Xuan Zheng and Zheng Fang and Yin Liu and Hanyu Zhang and Lihua Li and Qing Ye},
  title = {{Evaluation of the corneal topography based on deep learning}},
  booktitle = {Frontiers in Medicine},
  year = {2024},
  doi = {10.3389/fmed.2023.1264659},
}
Figure