The first version of the Retinal IMage database for Optic Nerve Evaluation (RIM-ONE) was published in 2011. This was followed by two more, turning it into one of the most cited public retinography databases for evaluating glaucoma. Although it was initially intended to be a database with reference images for segmenting the optic disc, in recent years we have observed that its use has been more oriented toward training and testing deep learning models. The recent REFUGE challenge laid out some criteria that a set of images of these characteristics must satisfy to be used as a standard reference for validating deep learning methods that rely on the use of these data. This, combined with the certain confusion and even improper use observed in some cases of the three versions published, led us to consider revising and combining them into a new, publicly available version called RIM-ONE DL (RIM-ONE for Deep Learning). This paper describes this set of images, consisting of 313 retinographies from normal subjects and 172 retinographies from patients with glaucoma. All of these images have been assessed by two experts and include a manual segmentation of the disc and cup. It also describes an evaluation benchmark with different models of well-known convolutional neural networks.

Deep learning systems, especially in critical fields like medicine, suffer from a significant drawback, their black box nature, which lacks mechanisms for explaining or interpreting their decisions. In this regard, our research aims to evaluate the use of surrogate models for interpreting convolutional neural network (CNN) decisions in glaucoma diagnosis. Our approach is novel in that we approximate the original model with an interpretable one and also change the input features, replacing pixels with tabular geometric features of the optic disc, cup, and neuroretinal rim. We trained CNNs with two types of images: original images of the optic nerve head and simplified images showing only the disc and cup contours on a uniform background. Decision trees were used as surrogate models due to their simplicity and visualization properties, while saliency maps were calculated for some images for comparison. The experiments carried out with 1271 images of healthy subjects and 721 images of glaucomatous eyes demonstrate that decision trees can closely approximate the predictions of neural networks trained on simplified contour images, with R-squared values near 0.9 for VGG19, Resnet50, InceptionV3 and Xception architectures. Saliency maps proved difficult to interpret and showed inconsistent results across architectures, in contrast to the decision trees. Additionally, some decision trees trained as surrogate models outperformed a decision tree trained on the actual outcomes without surrogation. Decision trees may be a more interpretable alternative to saliency methods. Moreover, the fact that we matched the performance of a decision tree without surrogation to that obtained by decision trees using knowledge distillation from neural networks is a great advantage since decision trees are inherently interpretable. Therefore, based on our findings, we think this approach would be the most recommendable choice for specialists as a diagnostic tool.

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.

The automatic location of the fovea is very useful for diagnosing retinal diseases. It is a complex problem for which different solutions have been proposed based on classical image processing and Deep Learning techniques. The method presented in this paper is based on histograms that combine spatial and color information in such a way that the spatial coordinates are incorporated into conventional color histograms as an additional dimension. The binarization of these histograms retains a considerable amount of relevant information from the original image, allowing them to be processed as if they were ordinary images. This approach to the problem results in a simple, fast and effective method for locating the fovea. Different experiments have been carried out with three popular sets of images: Messidor, REFUGE1 and DIARETDB1, and a comparison was made with other state-of-the-art techniques. Our results show that the proposed method, despite its simplicity, is capable of surpassing many of these techniques.

This paper systematically evaluates saliency methods as explainability tools for convolutional neural networks trained to diagnose glaucoma using simplified eye fundus images that contain only disc and cup outlines. These simplified images, a methodological novelty, were used to relate features highlighted in the saliency maps to the geometrical clues that experts consider in glaucoma diagnosis. Despite their simplicity, these images retained sufficient information for accurate classification, with balanced accuracies ranging from 0.8331 to 0.8890, compared to 0.8090 to 0.9203 for networks trained on the original images. The study used a dataset of 606 images, along with RIM-ONE DL and REFUGE datasets, and explored nine saliency methods. A discretization algorithm was applied to reduce noise and compute normalized attribution values for standard eye fundus sectors. Consistent with other medical imaging studies, significant variability was found in the attribution maps, influenced by the method, model, or architecture, and often deviating from typical sectors experts examine. However, globally, the results were relatively stable, with a strong correlation of 0.9289 (p < 0.001) between relevant sectors in our dataset and RIM-ONE DL, and 0.7806 (p < 0.001) for REFUGE. The findings suggest caution when using saliency methods in critical fields like medicine. These methods may be more suitable for broad image relevance interpretation rather than assessing individual cases, where results are highly sensitive to methodological choices. Moreover, the regions identified by the networks do not consistently align with established medical criteria for disease severity.

This work aims to leverage Shapley values to explain the decisions of convolutional neural networks trained to predict glaucoma. Although Shapley values offer a mathematically sound approach rooted in game theory, they require evaluating all possible combinations of features, which can be computationally intensive. To address this challenge, we introduce a novel strategy that discretizes the input by dividing the image into standard regions or sectors of interest, significantly reducing the number of features while maintaining clinical relevance. Moreover, applying Shapley values in a machine learning context necessitates the ability to selectively exclude features to evaluate their combinations. To achieve this, we propose a method involving the occlusion of specific sectors and re-training only the non-convolutional portion of the models. Despite achieving strong predictive performance, our findings reveal limited alignment with medical expectations, particularly the unexpected dominance of the background sector in the model’s decision-making process. This highlights potential concerns regarding the interpretability of convolutional neural network-based glaucoma diagnostics.