Selective fusion of structural and functional data for improved glaucoma detection

Abstract

This work proposes novel selective feature fusion of structural and functional data for improved glaucoma detection. The structural data such as retinal nerve fiber layer (RNFL) thickness measurement acquired by scanning laser polarimetry (SLP) is fused with the functional visual field (VF) measurement recorded from the standard automated perimetry (SAP) test. The proposed selective feature fusion exploits correspondence between structural and functional data obtained over multiple sectors. The correlation coefficients for corresponding structural-function sector pairs are used as weights in subsequent feature selection. The sectors are ranked according to the correlation coefficients and the first four highly ranked sectors are retained. Following our prior work, fractal analysis (FA) features for both structural and functional data are obtained and fused for each selected sectors respectively. These fused FA features are then used for glaucoma detection. The novelty of this work stems from (i) locating structure-functional sectoral correspondence; (ii) selecting only a few interesting sector pairs using correlation coefficient between structure-function data; (iii) obtaining novel FA features from these pairs; and (iv) fusing these features for glaucoma detection. Such a method is distinctively different from other existing methods that exploit structure-function models in that structure-function sectoral correspondences have been weighted and, based on such weights, only portions of the sectors are retained for subsequent fusion and classification of structural and functional features. For statistical analysis of the glaucoma detection results, sensitivity, specificity and area under receiver operating characteristic curve (AUROC) are calculated. Performance comparison is obtained with those of existing feature-based techniques such as wavelet-Fourier analysis (WFA) and fast-Fourier analysis (FFA). Comparisons of AUROC values show that our novel selective feature fusion method for discrimination of glaucomatous and ocular normal patients slightly outperforms other existing techniques with AUROCs of 0.98, 0.98 and 0.99 for WFA, FFA and FA respectively.