In-Depth Evaluation of Saliency Maps for Interpreting Convolutional Neural Network Decisions in the Diagnosis of Glaucoma Based on Fundus Imaging.

Glaucoma, a leading cause of blindness, damages the optic nerve, making early diagnosis challenging due to no initial symptoms. Fundus eye images taken with a non-mydriatic retinograph help diagnose glaucoma by revealing structural changes, including the optic disc and cup. This research aims to thoroughly analyze saliency maps in interpreting convolutional neural network decisions for diagnosing glaucoma from fundus images. These maps highlight the most influential image regions guiding the network's decisions. Various network architectures were trained and tested on 739 optic nerve head images, with nine saliency methods used. Some other popular datasets were also used for further validation. The results reveal disparities among saliency maps, with some consensus between the folds corresponding to the same architecture. Concerning the significance of optic disc sectors, there is generally a lack of agreement with standard medical criteria. The background, nasal, and temporal sectors emerge as particularly influential for neural network decisions, showing a likelihood of being the most relevant ranging from 14.55% to 28.16% on average across all evaluated datasets. We can conclude that saliency maps are usually difficult to interpret and even the areas indicated as the most relevant can be very unintuitive. Therefore, its usefulness as an explanatory tool may be compromised, at least in problems such as the one addressed in this study, where the features defining the model prediction are generally not consistently reflected in relevant regions of the saliency maps, and they even cannot always be related to those used as medical standards.

Oculus-Spark perimetry compared with 3 procedures of glaucoma morphologic analysis (GDx, HRT, and OCT).

PURPOSE: To compare diagnostic capabilities and agreement between Oculus-Spark perimetry and 3 procedures of glaucoma morphologic analysis. METHODS: A total of 102 normal eyes and 104 consecutive eyes with suspected or confirmed glaucoma (1 eye per subject) were analyzed in a prospective observational case-control study, using Spark strategy (Oculus Easyfield Perimeter), Heidelberg retinal tomograph (HRT), Zeiss laser polarimetry (GDx), and Cirrus optical coherence tomography (OCT). RESULTS: Spark first phase lasted 37 seconds and all 4 phases 2:34 minutes. Specificities and sensitivities were as follows: Spark mean deviation (MD) first phase (95.1%, 85.6%), MD second and final phases (95.1%, 86.5%), GDx-nerve fiber indicator (95.1%, 57.4%), HRT-Reinhard Burk discriminant function (95.1%, 52.9%), HRT glaucoma probability score (95.1%, 71.2%), Cirrus OCT vertical cup/disc ratio (96.1%, 85.6%), and Cirrus OCT retinal nerve fiber layer thickness (95.1%, 68.0%). Diagnostic agreement between second and final Spark MD phases was kappa=0.92; between phase 1 Spark/MD and Cirrus OCT/vertical C/D ratio was kappa=0.78. Agreements between the 2 Cirrus OCT indices was kappa=0.69 and between the 2 HRT indices was kappa=0.559. The correlation coefficient between second and final MD and PSD was 0.99, and 0.98 between the number of scotomatous points. There was high concordance in scotoma position in both phases (kappa=0.86). The linear correlation coefficients between the morphologic indices were 0.48-0.78, and between morphologic and functional indices 0.51-0.76. Correlation coefficients comparing morphologic and functional indices were similar in the first and the last phase (p>0.05 in all cases). CONCLUSIONS: Spark perimetry appears to show useful sensitivity and specificity, even in the first phase, and good agreement with the morphology.

Linear regression analysis of the cumulative defect curve by sectors and other criteria of glaucomatous visual field progression.

PURPOSE: To analyze the progression of visual field loss using sector analysis of the cumulative defect curve and other procedures. METHODS: Visual fields of 260 glaucomatous eyes were analyzed over 2.8 years (SD = 1.2) with at least five examinations (mean 6.9; SD = 2.0) using Octopus 311 perimeter and TOP strategy. The authors applied Threshold Noiseless Trend (TNT) program, which performs local filtering of threshold to reduce fluctuation, and analyzed five criteria: 1) a score based on significant progression of eight sectors of the cumulative defect curve (CD); 2) a score based on the presence of points (PO) with significant progression; 3) global progression (GL) of all local deviations; 4) progression of mean defect (MD); 5) progression of the square root of loss variance (sLV). The authors estimated false diagnoses (FD), randomly reordering examinations of each patient. An index of focality of progression (FI) was obtained. RESULTS: sLV presented low sensitivity and GL low specificity. CD and PO presented twice the sensitivity of MD, often proving earlier indicators. The authors observed significant progression of some of the three criteria in 17.5% of the cases when MD <6 dB and in 20.7% when MD >6 dB (FD=5.7%). Agreement between two criteria occurred in 6.8% of cases with MD <6 dB and in 11.6% when MD >6 dB (FD=1.9%). Result reproducibility in successive examinations was observed in 9.9% of cases (FD=1.3%). Focality of progression increased with MD. CONCLUSIONS: PO and CD indicate suspected progression earlier than MD. Reproduction of results in successive examinations or agreement between criteria allows confirmation of progression.

Stabilization and comparison of TOP and Bracketing perimetric strategies using a threshold spatial filter.

BACKGROUND: To evaluate a new perimetric spatial filter that takes into account relations of dependence between regions of the glaucomatous visual field. METHODS: 51 glaucoma patients and 30 controls were examined using the Octopus 1-2-3 on four occasions using program 32; two with TOP and two with Bracketing (BRA) strategy. Each threshold was replaced by a filtered threshold, calculated as the mean of its own value and the four points best correlated with it, weighted with the correlation coefficient (r) that relates them. RESULTS: Application of the filter had minimal effect on the absolute mean defect (MD) but reduced the square root of loss variance (sLV) by 17.6% in TOP and 28.8% in BRA, increasing the similarity between their results. Filtered TOP and BRA thresholds were more similar than those obtained in the two unfiltered BRA examinations. Filtering reduced the value of short fluctuation by 28.6% in TOP and 45.4% in BRA and reduced sLV fluctuation by 14.3% in TOP and 24.2% in BRA, thus harmonizing the two strategies for these two parameters (p > 0.05). The influence of filtering on MD fluctuation was minimal (p > 0.05). In normal subjects the number of points beyond 5 dB of normality (not reproducible false scotomas) reduced from 1.5% to 0.2% in TOP and from 4.9% to 1.6% in BRA, while in glaucoma patients changes were minimal. CONCLUSIONS: The proposed spatial filter stabilized perimetric results, acting with greater effect on BRA than on TOP, making their results more similar.