Editorial: Efficient artificial intelligence (AI) in ophthalmic imaging

Yanda Meng, Meng Wang, Haoyu Chen, Zheng YalinFrontiers in Medicine 2024 · DOI · OpenAlex

Abstract

Recent advancements in deep learning have catalyzed the development of automated frameworks for ophthalmic diagnostics. Wang et al. (2023) proposed a novel system to extract twelve three-dimensional parameters from automatically segmented hyperreflective foci (HRF) in optical coherence tomography (OCT), a critical task for understanding various ocular diseases. Furthermore, retinal vessels serve as vital biomarkers for conditions such as hypertensive retinopathy, though manual identification remains resource-intensive. Beyond vessel analysis, axial length (AL) has emerged as a significant parameter defining refractive status and its strong association with macular complications. This editorial synthesizes research focused on "efficient AI"—models that are computationally practical and require less manual labeling while maintaining high diagnostic precision.

Introduction

Ophthalmic imaging modalities, including Optical Coherence Tomography (OCT), fundus photography, and Fluorescein Angiography (FFA), generate vast quantities of visual data essential for tracking diseases like age-related macular degeneration and glaucoma. However, the manual analysis of these images represents a significant clinical bottleneck. It is labor-intensive, prone to human error, and characterized by high inter-observer variability. The core challenge addressed in this research collection is the transition toward efficient AI: models that reduce the computational "tax" and the need for expensive manual labeling while performing complex tasks like 3D parameter extraction and multi-view image registration.

Methodological Innovations

The featured research introduces several novel algorithmic approaches designed to optimize diagnostic workflows:

  • Advanced Attention Mechanisms: Researchers introduced the Neighbored Attention U-Net (NAU-Net) to balance high-performance segmentation with lower computational costs. Additionally, a heterogeneous feature cross-attention neural network was developed specifically for retinal vessel segmentation.
  • Unsupervised and Weakly Supervised Learning: To mitigate the burden of data labeling, a Generalized Polynomial Transformation (GPT) model was created for unsupervised fundus image registration. Similarly, a weakly supervised model was successfully deployed to localize retinal detachment.
  • Hybrid Architectures: For ocular surface disease classification, the DBPF-Net utilizes a "conformer" backbone—combining Transformers and CNNs—to achieve accurate four-way classification of conditions like pterygium.
  • Generative AI and LLMs: The integration of Large Language Models (LLMs) is explored as a tool to assist clinicians in information retrieval and patient communication, moving AI beyond image analysis into the realm of clinical decision support.

Key Insights and Clinical Results

The shift toward efficiency has yielded several high-impact results across diverse datasets, including one study utilizing over 24,000 ultra-widefield fundus images:

  1. Automated 3D Reconstruction: The successful extraction of 3D parameters from segmented hyperreflective foci provides a more granular understanding of disease progression than traditional 2D analysis.
  2. Axial Length Estimation: Deep learning models have demonstrated the ability to estimate axial length directly from macular OCT images, even in eyes without maculopathy. This offers a non-invasive alternative to traditional biometry, crucial for managing myopia.
  3. Robust Localization: Weakly supervised models proved effective in outlining anatomical regions of retinal detachment, even in emergency-style imaging scenarios where high-quality labeling is often unavailable.
  4. Efficiency over Brute Force: The success of the GPT model for image registration suggests that high-quality alignment can be achieved without perfectly labeled ground-truth pairs, significantly lowering the barrier for clinical implementation.

Future Directions and Limitations

While these advancements are promising, the editorial identifies several hurdles for the next generation of AI tools. The rapid pace of imaging hardware development requires models to be continuously updated to maintain compatibility. Furthermore, moving from technical success to daily clinical practice requires navigating complex regulatory landscapes and ensuring models generalize across diverse patient populations and different device manufacturers. The ultimate goal remains a faster, more accurate, and more accessible AI-assisted ophthalmic practice.

Summary updated

Abstract

Wang et al. (2023) proposed a new automated framework driven by recent advances in deep learning 13 to extract 12 three-dimensional parameters from the automatically segmented hyperreflective foci in 14 optical coherence tomography (OCT). This is because the fast and automated reconstruction of retinal 15 hyperreflective foci (HRF) is of great importance for understanding many eye-related diseases.Retinal vessels play a pivotal role as biomarkers in detecting retinal diseases, including hypertensive 17 retinopathy. The manual identification of these retinal vessels is both resource-intensive and time- Axial length (AL) is a widely discussed parameter, significant not only for defining the eye's refractive 44 status but also due to its strong association with retinal and macular complications. The excessive
BibTeX
@inproceedings{meng2024editorial,
  author = {Yanda Meng and Meng Wang and Haoyu Chen and Zheng Yalin},
  title = {{Editorial: Efficient artificial intelligence (AI) in ophthalmic imaging}},
  booktitle = {Frontiers in Medicine},
  year = {2024},
  doi = {10.3389/fmed.2024.1523647},
}
Figure