蜜芽app

Rates of age-related macular degeneration (AMD), a leading cause of permanent vision loss, continue to rise as the population ages. In patients with progressing AMD, neovascularization can occur, leaving them with the devastating effect of blindness caused by fluid accumulation and atrophy within the retina.¹

While treatment has advanced with the use of vascular endothelial growth factor inhibitors (anti-VEGF), real-world outcomes have lagged due to undertreatment, possibly from poor detection of retinal fluid.² Studies to understand the disease have relied heavily on measurements of central retinal subfield thickness (CSFT) rather than the amount of accumulating retinal fluid. As OCT technology advanced, intraretinal and subretinal fluid has been more easily detected, and is now known to correlate better with visual function than CSFT.^3,4 

The authors of a new study published in Eye believe 鈥渁n innovative approach to precisely and quantitatively measure clinically relevant features such as fluid in an automated fast and objective manner is needed to support clinicians in making reliable treatment decisions.鈥�¹

In their trial, the investigators sought to understand how well automated fluid detection could identify retinal fluid activity in optical coherence tomography (OCT) scans, by correlating human expert grading and automated methods with CSFT and fluid volume values. To be eligible for the study, patients had to have been treated with anti-VEGF therapy for neovascular AMD (nAMD) in the previously conducted HAWK and HARRIER clinical trials assessing the efficacy of brocilizumab for nAMD.

What the study focused on

To assess the accuracy between deep learning-based algorithms and measurements from a certified reading center, the study used 44,903 spectral domain OCT (SD-OCT) volume scans from 2771 patients with nAMD enrolled in the prospective, multicenter HAWK and HARRIER randomized clinical trials. After applying exclusion criteria, 23,986 and 17,161 OTC scans were included from the HAWK and HARRIER studies, respectively.

The HAWK and HARRIER studies demonstrated a moderate and low correlation with CSFT at baseline for intraretinal fluid (IRF) and subretinal fluid (SRF) volumes, respectively:

• IRF: HAWK, r=0.54; HARRIER, r=0.62
• SRF: HAWK, r=0.29; HARRIER, r=0.22

When patients were under anti-VEGF therapy, the CSFT correlations between IRF and SRF remained low:

• IRF: HAWK, r=0.44; HARRIER, r=0.34
• SRF: HAWK, r=0.38; HARRIER, r=0.45

鈥淯nder treatment, patients lost 89.9% to 93.2% of mean IRF and 81.7% to 84.2% of mean SRF compared to baseline, with only tiny amounts of fluid remaining in the retina,鈥� the authors wrote.¹

Deep learning found to be effective

Based on the results of their study, the authors concluded that 鈥淸d]eep learning-based image evaluation is able to bridge this gap using automated algorithms convincingly,鈥�¹ the 鈥済ap鈥� being in reference to reliable data regarding retinal thickness and fluid amounts used in the treatment decision-making process.

In regard to IRF, which the authors cite as being 鈥渢he most important indicator for visual acuity correlations over as long as 5 years,鈥�¹ area under the curve (AUC) values for the detection of fluid in the central millimeter between expert grading and algorithmic reading were 0.93 in HARRIER and 0.85 in HAWK. AUC values for the detection of fluid for SRF were both 0.87 in the central millimeter between the 2 methods.

In addition, the authors also assessed AUC values for the central 3 mm and the central 6 mm:

• 3 mm IRF: 0.92 for HARRIER, 0.86 for HAWK
• 6 mm IRF: 0.90 for HARRIER, 0.85 for HAWK
• 3 mm SRF: 0.90 for HARRIER, 0.91 for HAWK
• 6 mm SRF: 0.90 for HARRIER, 0.91 for HAWK 

Limitations and conclusions

Because these deep-learning tools are a new advancement, there are no prospective trials that have thoroughly assessed their validity. Still, the current study鈥檚 retrospective design is considered a limitation. Further, the algorithm was created and validated using 3 major OCT scanning devices鈥擲pectralis, Cirrus, and Bioptigen鈥攂ut not with a fourth, Topcon. Topcon devices contributed less than 5% of the total data used in the study.

Nevertheless, given the reliability found with automated deep learning in this study, the authors wrote that these data show 鈥淐SFT values are weak indicators for fluid activity in nAMD.鈥� They also highlighted the opportunity their learnings provide for the future of treatment for patients with nAMD, as large volumes of data can be analyzed and utilized at point of care to improve real-world outcomes.

鈥淎utomated quantification of fluid types,鈥� they concluded, 鈥渉ighlight the potential of deep learning-based approaches to objectively monitor anti-VEGF therapy.鈥�¹

Published: October 24, 2023