A recent study published in Transitional Vision Science & Technology utilized the artificial intelligence (AI) method Deep Approximation of Retinal Traits (DART) to investigate the possible use of fractal dimension (FD) as a biomarker for systemic disease.
Give me some background first.
As shown in previous reports, AI has emerged as a valuable tool in the eye care industry and this study is no exception.
DART is an artificial intelligence-based method for computing FD.
Unlike traditional methods for computing FD such as VAMPIRE, which requires high-quality images to produce calculations, DART utilizes a deep learning model trained to ignore variations in image quality and take into account all available information in an image.
This has made DART more robust than traditional approaches.
Now, talk about the study.
This study examined clinical records sourced from eye examinations of a mixed-age, primary-care population that occurred between 2017 and 2022 at the Vision Centre at Glasgow Caledonian University (GCU).
Record cards were completed for each patient during their examination as well as fundus images. Retinal images were taken using swept-source optical computed tomography.
The following patient information was collected:
- Age at visit
- Sex
- General health status (collected from the history and symptoms field in the record card)
Note: No images were excluded due to poor quality in order to avoid bias and evaluate the effectiveness of FD under real-world conditions.
Who was included in the study?
A total of 96 participants (183 eyes) with the following demographics:
- Aged 18 to 8
- 58 females (60%); 38 males (40%)
- Median age of 61.80 years (interquartile range [IQR], 33.20)
- youngest patients aged 18 and the oldest aged 81
The researchers had access to retinal images of 183 patients. In nine cases, only a single image was available, which they noted as indicating that only one eye was examined.
The most recent visit of each individual and all images available were used for analysis.
Findings?
The investigators found that prevalent systemic health conditions were associated with a significant decrease in FD. They also found that while FD decreased with age, it did not substantially decrease for all individuals.
How did age impact results?
The researchers discovered:
- Age was associated with lower FD (P < 0.001 univariate and when adjusting for image quality).
- FD variance was higher in older patients
- Some patients aged 60+ had FD similar to patients in their 20s.
- Prevalent systemic conditions were significantly associated with lower FD after adjusting for image quality and age (P = 0.037)
How did DART perform?
The researchers noted that despite DART being designed to be more robust, there was still a significant association (P < 0.001) between image quality and FD.
Limitations?
The study authors stated one limitation of their study was participants may not have mentioned prevalent conditions, were not deemed relevant for their record, or may have been undiagnosed at the time of their examination.
Further, there was a lack of information regarding the severity and duration of patients’ conditions.
Go on …
Another limitation was the coarseness of the information about prevalent systemic conditions and a lack of information concerning incident systemic conditions. Additionally, only a single quality annotator was used.
Lastly, the researchers stated that while DART is more robust to image quality when compared to traditional methods, some of their images may have been too poor quality.
However, they added that even the worst-quality images provided enough information for their FD estimates.
Expert opinion?
The study authors stated that while their results were promising, “future work should also more closely investigate what specific vascular changes are captured by FD.”
Take home.
According to the researchers, FD has potential use in regular screening settings. While DART is more robust than traditional methods for computing FD, the question regarding the impact of image quality on results remains.