Previous studies focused on detecting one type of pathological fluid, mostly IRF causing macular edema.
9,11,12,17 Lee et al.
12 used a U-Net architecture to detect IRF with an average Dice score of 0.73. Contrary to our approach, the network operated on image patches instead of the entire B-scans, and the total prediction time was over 13 seconds per B-scan compared to 0.04 second in our case. Venhuizen et al.
11 showed that adding retinal segmentation prior helped with detecting intraretinal cystoid fluid, achieving an overall Dice score of 0.75. Roy et al.
18 and Asgari et al.
17 also showed that adding additional prior information in the form of retinal layers helped segmenting fluids in drusen and diabetic retinopathy. This corresponds well with our observation, and it is indeed not surprising because fluid definitions are closely related to the retinal layers. The detection of IRF and SRF in different retinal diseases was investigated by Schlegl et al.,
10 where the model obtained a precision of 0.78 and 0.81, and a sensitivity of 0.63 and 0.71, for IRF and SRF, respectively. Furthermore, they compared the distributions of predicted and annotated fluid volumes achieving an
R2 of 0.68 and 0.65 for IRF and SRF in Spectralis scans, and Pearson coefficients of 0.86 and 0.85, respectively. The multiclass fluid detector introduced by Lu et al.
8 and Lee et al.
9 are the closest to our method, because they distinguished simultaneously between IRF, SRF, and PED, obtaining Dice scores between 0.74 to 0.85 and 0.75 to 0.86, respectively. However, the method of Lu et al.
8 consisted of several refining steps, which increased the computational burden and required more parameter tuning. On the contrary, our method is trained in a single step, and it greatly reduces the processing time and complexity. In addition, these publications did not contain information about fluid volumes and their correlations.