Purpose:
To automatically estimate the total fluid volume in 3D OCT scans of patients with CNV due to AMD. The total fluid volume is important in the determination of the CNV treatment regimen as well as for assessing the effects of the treatment.
Methods:
For this work 15 independent, macular-centered SD-OCT scans (Zeiss Cirrus HD-OCT) from 15 eyes of 15 subjects with CNV due to AMD were acquired. Each SD-OCT scan consisted of 200×200×1024 voxels with a voxel size of 30×30×2µm. The retinal layers in each SD-OCT scan were automatically segmented. We then applied a previously developed technique to find the location of the fluid in the XY plane, the footprint of the fluid filled region. Using this footprint, the layer segmentation was adapted to better estimate the bottom surface of the visible retina in the 3D scan. This bottom surface is often deformed due to the pressure exerted on it by the fluid. A large set of 3D structural and textural features was extracted for each voxel in the images. These included wavelet based texture features and features based on the thickness of the layers from the layer segmentation. The reference standard used for training and evaluating the method was generated by a retinal specialist who marked the voxels in each of the scans that were inside a fluid filled region. Using a specially developed tool, this process took approximately 1 hour per scan. Once the reference standard was available a statistical classifier was trained to assign each voxel in a 3D OCT scan a likelihood between 0 and 1 that the voxel is part of a fluid-filled region. By thresholding this value, an estimated fluid volume was calculated. The optimal threshold was determined on the training set.
Results:
A leave-one-scan-out experiment was conducted where the system was trained 15 times, each time leaving out a single scan for which the system then estimated the fluid volume. The correlation between the estimated fluid volume and the reference standard fluid volume was calculated to be 0.94 with a single outlier and 0.97 if the outlier was removed (see Figure 1).
Conclusions:
We have shown, for the first time, that an automated method can estimate the total fluid volume in a 3D OCT scan of a patient with CNV due to AMD.
Keywords: image processing • imaging/image analysis: clinical • retina