Abstract
Purpose: :
To detect localized glaucomatous changes with type I error control using the false discovery rate control (FDR) method within the POD framework. FDR controls only the mean false positive rate, i.e. the actual false positive rate achieved may be greater than the desired rate (Balasubramanian M, et al, IOVS 2009; 50: ARVO E-Abstract 2573). Therefore, we also evaluate Bonferroni correction (BC) and generalized family-wise error rate control (gFWER; Lehmann E & Romano J, Annals of Statistics 2005) methods that provide a strict upper bound for the actual false positive rate.
Methods: :
For each eye, a POD baseline subspace was constructed and baseline subspace representations of each follow-up were generated (Balasubramanian M, et al, IOVS 2010). At each superpixel (4×4 pixels), the significance of retinal height change between each follow-up and its baseline subspace representation was estimated using a 2-factor fixed-effects ANOVA model. To identify locations with significant reduction in retinal height (red-pixels) and to construct POD change significance maps, p-value cutoffs were estimated at a false positive rate of 5% using FDR, BC, and gFWER. Observed positive rates (OPR) were calculated for each follow-up as the ratio of total no. of red-pixels within the optic disk to disk size. Progression was defined as the presence of one or more follow-ups with OPR > 5% for FDR and gFWER and OPR > 0% for BC. 167 participants (246 eyes) with ≥4 HRT-II exams from the UCSD Diagnostic Innovations in Glaucoma Study were included; 36 eyes progressed by stereophotographs or showed likely progression on SAP Guided Progression Analysis (progressors). All other eyes (n = 210) were considered to be non-progressing. Specificities were estimated in 21 longitudinal normal eyes (normals).
Results: :
Sensitivities in progressors: FDR = 69%; BC = 100%; gFWER = 86%. Specificities in non-progressors: FDR = 49%; BC = 0%; gFWER = 34%. Specificities in normals: FDR = 95%; BC = 0%; gFWER = 81%.
Conclusions: :
POD change significance maps are useful for visual inspection of retinal locations with likely glaucomatous changes and for deriving objective criteria for glaucoma progression with reduced need for confirmatory follow-ups. BC was anti-conservative in detecting progression. gFWER improved over BC, but had low specificity in non-progressors. The FDR based progression criteria provided an optimal diagnostic accuracy with moderate sensitivity in progressors, high specificity in normals and moderate specificity in non-progressors.
Keywords: imaging/image analysis: clinical • optic disc • computational modeling