Individualized deep learning auto-encoder (DL-AE)-based OCT deviation maps for improved glaucoma progression detection

Abstract

Purpose : To compare individualized, eye-specific Spectralis OCT regional RNFL change detection maps developed using unsupervised DL-AE strategies to circumpapillary RNFL thickness change over time for detection of glaucomatous progression.

Methods : Forty-four progressing glaucoma eyes (by stereophotograph assessment),189 non-progressing glaucoma eyes (by stereophotograph assessment) and 109 stable healthy eyes from the Diagnostic Imaging in Glaucoma Study (DIGS) and the African Descent and Glaucoma Evaluation Study (ADAGES) were followed over 3 to 5 years with 4 to 10 visits using Spectralis OCT. Fifty stable glaucoma eyes (tested weekly for five weeks) were used to train DL-AEs to identify regional change greater than measurement variability. The San Diego Automated Layer Segmentation Algorithm (SALSA) was used to automatically segment the RNFL layer from raw 3-D OCT images. For each longitudinal series, DL-AEs were used to generate individualized eye-based change detection maps by identifying RNFL regions of likely progression and no change. Sensitivities and specificities for detecting regional change over time and rates of change over time were compared for the DL-AE model and circumpapillary RNFL circular thickness maps (obtained within a 2.22 mm to 3.45 mm annulus centered on the optic nerve).

Results : Sensitivity for detecting change in progressing glaucoma eyes was greater for DL-AE maps than RNFL circular thickness maps (0.90 and 0.63, respectively), while specificity for detecting non-progression in non-progressing glaucoma eyes was similar (0.92 and 0.93, respectively); 40% more progressing eyes where identified using DL-AE maps compared to RNFL circular thickness maps. Mean (95% CI) rates of change in DL-AE map regions of likely progression were significantly faster than for RNFL circular thickness maps in progressing glaucoma eyes [-1.28 µm/yr (-1.38 µm/yr, -1.15 µm/yr) vs. -0.83 µm/yr (-0.93 µm/yr, -0.72 µm/yr)], non-progressing glaucoma eyes [-1.03 µm/yr (-1.21 µm/yr, -0.93 µm/yr) vs. -0.78 µm/yr (-0.88 µm/yr, -0.67 µm/yr)] and healthy eyes [-0.83 µm/yr (-0.92 µm/yr, -0.73 µm/yr) vs. -0.65 µm/yr (-0.74 µm/yr, -0.56 µm/yr)].

Conclusions : By tailoring analysis to the individual patient, individualized regions of interest identified using unsupervised deep learning auto-encoder analysis of OCT images show promise for improving assessment of glaucomatous progression.

This is a 2020 ARVO Annual Meeting abstract.