Multi-layer 3D Simultaneous Retinal OCT Layer Segmentation: Just-Enough Interaction for Routine Clinical Use

All current fully automated retinal layer segmentation methods fail in some subset of clinical 3D Optical Coherence Tomography (OCT) datasets, especially in the presence of appearance-modifying retinal diseases like Age-related Macular Degeneration (AMD), Diabetic Macular Edema (DME), and others. In the presence of local or regional failures, the only current remedy is to edit the obtained segmentation in a slice-by-slice manner. This is a very tedious and time-demanding process, which prevents the use of quantitative retinal image analysis in clinical setting. In turn, the non-existence of reliable retinal layer segmentation methods substantially limits the use of precision medicine concepts in retinal-disease applications of clinical ophthalmology. We report a new non-trivial extension of our previously-reported LOGISMOS-based simultaneous multi-layer 3D segmentation of retinal OCT images. In this new approach, automated segmentation of up to 9 retinal layers defined by 10 surfaces is followed by visual inspection of the segmentation results and by employment of minimally-interactive correction steps that invariably lead to successful segmentation thus yielding reliable quantification. The novel aspect of this "Just-Enough Interaction" (JEI) approach for retinal OCT relies on a 2-stage coarse-to-fine segmentation strategy during which the operator interacts with the LOGISMOS graph-based segmentation algorithm by suggesting desired but approximate locations of the layer surfaces in 3D rather than performing manual slice-by-slice corrections. As a result, the efficiency of the reliable analysis has been improved dramatically with more than 10-fold speedup compared to the traditional retracing approaches. In an initial testing set of 40 3D OCT datasets from glaucoma, AMD, DME, and normal subjects, clinically accurate segmentation was achieved in all analyzed cases after 5.3 +/- 1.4min/case devoted to JEI modifications. We estimate that reaching the same performance using slice-by-slice editing in the regions of local segmentation failures would require at least 60 min of expert-operator time for the 9 segmented retinal layers. Our JEI-LOGISMOS approach to segmentation of retinal 3D OCT images is now employed in a larger clinical-research study to determine its usability on a larger sample of OCT image data.