Advances in AI stand to greatly improve diagnosis, treatment of AMD, but hurdles remain
Key takeaways:
- AI in combination with OCT and fundus photography helped to diagnose and treat AMD in a variety of studies.
- AI’s efficacy is limited by insufficient or unbalanced data and limited generalizability.
Although AI shows significant promise in diagnosing and managing age-related macular degeneration, its efficacy is reduced by bias, data quality and limited generalizability, according to a review published in BMJ Open Ophthalmology.
Health professionals face obstacles regarding early diagnosis of AMD as well as customizing treatment and monitoring disease progression. However, the ability of AI to analyze retinal image datasets can improve AMD diagnosis and screening, according to Yundi Gao, of Nanchang University and Beijing Bright Eye Hospital in China, and colleagues.
“AI-driven retinal diagnostic systems also provide personalized treatment plans tailored to each patient’s characteristics and condition,” they added. “By analyzing retinal images and clinical data, these systems can predict disease progression and recommend appropriate treatments.”
The expanded use of this technology inspired the researchers to conduct a systematic review to highlight the benefits of using AI as a diagnostic and treatment tool for AMD and how AI has progressed over the past 2 decades in this role.
The researchers searched the Web of Science database to identify 13 studies published between August 2005 and March 2024 involving AI, AMD and a form of retinal data imaging (fundus photography, OCT or fluorescein fundus angiography).
AI in fundus photography
First, the researchers analyzed AI applied to AMD based on fundus photography and the use of machine learning models for retinal image analysis.
According to their review, they found that machine learning models were noninferior to human observers in automated AMD risk assessment, which includes prediagnosis of the disease and an analysis of risk progression.
Further, deep learning models, which process images directly thus minimizing errors due to feature computation and segmentation, in combination with fundus photography were comparable to experts in identifying AMD at a reduced cost and with greater efficacy. For example, the authors reviewed a study that featured a deep learning model that distinguished AMD subtypes and achieved classification accuracies of more than 90% for dry AMD and wet AMD.
AI applied to OCT images
Next, the researchers found that deep learning AI models effectively classify high-resolution OCT images to detect AMD. Notably, researchers have progressed from using AI to simply distinguish AMD images from normal images to tracking the stages of the disease and differentiating AMD from other macular diseases, allowing for more rapid and accurate diagnoses and enabling timely interventions and treatment.
Also, the authors found that the rise of deep learning has enabled the development of advanced algorithms for OCT image segmentation, such as CNN and U-Net. These new algorithms show “significant promise” in performing segmentation tasks, which could improve the monitoring and understanding of AMD progression and pathology, the authors wrote. They noted that one novel deep learning model, AR U-Net++, was able to identify the exact location and depth of retinal fluid in the retinal layers.
The researchers also studied AI’s ability to use OCT images to predict AMD progression and treatment efficacy. They found that AI allows clinicians to more accurately and quickly determine AMD severity and progression by analyzing retinal thickness and pathological traits. Further, AI is particularly helpful in evaluating patient eligibility for anti-VEGF medications as a treatment for wet AMD, considering that some patients respond poorly to those medications.
Finally, the researchers found that combining different types of imaging data in multimodal models can improve diagnostic and predictive accuracy.
Limitations, next steps
The authors noted several limitations to the efficacy of AI in diagnosing and managing AMD. For one, current datasets are often biased, oversampling advanced cases and lacking inclusion of less common subtypes of cases. Because data so heavily impact model performance, data that are biased or of lower quality and consistency can harm the model’s accuracy, they noted.
Also, AI models have limited generalizability due to their dependence on specific datasets, which makes clinical implementation more difficult.
“The integration of AI in retinal disease management presents promising solutions to ongoing challenges in AMD,” Gao and colleagues wrote.
“Future research should prioritize collecting and annotating more OCT images, integrating additional imaging data to enhance model performance and improving model interpretability through visualization and attention mechanism,” they added.
Perspective
Back to Top
This study is focused on the application of AI tools such as machine learning and deep learning to enhance the diagnosis of AMD and predict more effectively its treatment outcomes. These AI techniques leverage advanced technologies, including fundus photographs, OCT and fluorescein fundus angiography.
The results of this study are based on the review of primary studies that investigated the application of AI techniques in the diagnosis, treatment and management of AMD. The review highlights the prospects of using AI techniques to diagnose, treat and monitor AMD and particularly their usefulness in identifying early cases of the disease and tailoring treatment to individual patients.
Using advanced AI models, high diagnostic accuracy, sensitivity and specificity for AMD have been observed with OCT images and other fundus imaging tools. However, it is unclear how the findings of this study can be applied in practical terms to improve the diagnosis and management of AMD.
Although the studies included in the review met clear eligibility criteria, the quality of the articles selected lacks consensus as to how they were selected and included as part of the analyses. Thus, caution should be applied in interpreting the findings of the study to clinical practice.
Collapse
Read more about