Performance Of Automated Retinal Fluid Identification From Oct Images In Real-World Management Of Neovascular Age-Related Macular Degeneration

Outcomes in neovascular age-related macular degeneration (nAMD) are often inferior in real-world practice despite substantial efforts. This is mostly due to a high variability in the assessment of OCT images obtained during long-term monitoring lacking precision and repeatability. Automated clinical decision support systems (CDSS) based on artificial intelligence (AI) are able to  objectively quantify macular fluid compartments such as intraretinal fluid (IRF), subretinal fluid (SRF) and pigment epithelium detachment (PED) in nAMD. This study analyzes the performance of the first CDSS on optical coherence tomography (OCT) data from a real-world outpatient clinic.

Optical coherence tomography (OCT) data from a real-world outpatient clinic in a tertiary referral center (Department of Ophthalmology at the Medical University of Vienna, Austria).

A multi-class deep learning-based method was developed to identify regions of SRF, IRF and PED in nAMD eyes. Dilated convolutions are used to detect fluid regions at multiple features and ensembling methods were used to increase confidence in the final pixel label. The dataset was comprised of 219 SD-OCT volumes  (Spectralis, Heidelberg Engineering, Germany), with 6210 b-scans from 155 different nAMD patients. Type 1, 2 and 3 nAMD from both treatment-naïve and therapeutic follow-ups were included. A combination of custom-built augmentation tools for shadowing, vignetting, noise, blur and affine transformation were utilised. The dice score is used as a primary metric.Confidence intervals were calculated using Wilson confidence interval.

Analyses contained a training set of 136 OCT volumes with 3620 annotated B-scans, the validation set contained 40 OCT volumes with 1314 annotated B-scans, and the test set contained 43 OCT volumes with 1276 annotated B-scans.  The Pearson correlations in the central 1mm for each fluid volume showed strong correlations between automated and the human expert gradings  with values for IRF: r=0.9998, p<0.001; SRF: r=0.9940, p<0.001, and PED: r=0.9848, p<0.001). Fluid detection was consistently well performed by the algorithm for each compartment on the B-scan level: IRF (dice: 0.9207, sensitivity: 0.8960, precision: 0.9468), SRF: (dice: 0.9660, sensitivity 0.9733, precision 0.9589), and PED (dice: 0.9153, sensitivity 0.9450, precision 0.8873). 

Disease activity and therapeutic response in nAMD can be measured in a precise and repeatable manner using AI-based CDSS such as the Vienna Fluid Monitor. Quantifications of macular fluids enable clinicians to base their re-treatment criteria on objective evaluation performed fast and in real-time in a busy clinic. With the use of accurate fluid quantification every clinician is empowered to high quality and transparency in AMD management introducing precision medicine and personalized treatment by saving costs and reducing workload.