Chimp-Inspired AI Hits 98.94% Accuracy in Glaucoma Detection From Fundus Images

Glaucoma, a condition in which elevated intraocular pressure gradually damages the optic nerve, remains one of the leading causes of irreversible blindness worldwide, and catching it early is often the difference between preserved and permanently lost vision. Standard clinical tests such as tonometry and pachymetry can detect the disease, but they are time consuming and heavily dependent on the skill of the examiner. A new study published in Neural Computing and Applications by Gaurav Pandey, Rajneesh Sharma, Amit Kukker, and Ronit Singh Negi proposes an automated framework that classifies stages of glaucoma from retinal fundus images, reaching 98.94 percent accuracy on the binary task of distinguishing healthy eyes from glaucomatous ones and 98.55 percent accuracy on the more demanding three-class task of separating normal, early, and advanced glaucoma.

The pipeline begins with a two-stage preprocessing scheme designed to suppress noise while preserving the fine structural cues that clinicians rely on. In the first stage, Bi-dimensional Empirical Mode Decomposition (BEMD) is applied to the fundus images to obtain Intrinsic Mode Functions, which decompose the image into oscillatory components adapted to the data itself rather than to a fixed basis. In the second stage, a Bi-dimensional Empirical Wavelet Transform (BEWT) follows, suppressing noise while retaining high-frequency information such as the sharp boundaries of the optic disc and the texture of the retinal nerve fiber layer. This hybrid empirical decomposition, the authors argue, extracts the most informative components of the image before any feature is computed, giving the downstream classifier a cleaner and more discriminative representation to work with.

From the processed images, the team computes a set of discriminative descriptors including Energy, Fractal Dimension, and several entropy measures. These features quantify the textural complexity and information content of the retina, which are known to change as glaucoma progresses. ReliefF, a filter-based feature selection algorithm, then picks out the most informative features, Linear Discriminant Analysis (LDA) compresses the feature space into a lower-dimensional representation, and Fisher’s index is used to rank the features prior to classification. This layered approach to feature engineering ensures that only the most relevant descriptors reach the classifier, reducing computational burden and guarding against overfitting.

The heart of the framework is a novel classifier the authors call Chimp-Optimized Fuzzy Q-Learning, or ChOAFQL. Fuzzy Q-learning combines reinforcement learning, in which an agent learns by receiving rewards for good decisions, with fuzzy inference systems, which handle uncertainty through graded membership rather than crisp categories. The chimp optimization algorithm, a nature-inspired metaheuristic that mimics the social hierarchy and foraging behavior of chimpanzees, is used to tune the parameters of the fuzzy Q-learning system. The result is a classifier that adapts its decision boundaries to the data with unusual flexibility, outperforming standard Fuzzy Q-learning, Genetic Algorithm-based Fuzzy Q-learning, and a range of benchmark classifiers across both datasets.

To verify that these results are not tied to a particular data split, the researchers ran 5-fold cross-validation, varied the train-test ratios across 70:30, 75:25, and 80:20 configurations, and stress-tested the model on images corrupted with Gaussian blur and downscaling. Performance remained consistent under all of these conditions, suggesting that the framework is robust to the kinds of image degradation that occur in real clinical settings, where fundus photographs vary widely in quality and resolution. Pairwise McNemar’s tests further confirmed that the improvements over FQL and GAFQL are statistically significant, with p-values below 0.05 on both datasets.

The evaluation was carried out on publicly available datasets, including RIM-ONE and Drishti-GS, two of the most widely used open retinal image databases for optic nerve assessment. The use of public data strengthens the reproducibility of the results and allows other research groups to benchmark their own methods against the same images. The authors note that no funds, grants, or other support were received during the preparation of the manuscript, and that they have no known competing financial interests or personal relationships that could have influenced the work.

What makes this approach particularly interesting from a technical standpoint is the way it bridges classical signal processing and modern computational intelligence. Rather than relying on a single monolithic deep network, the framework decomposes the problem into interpretable stages: empirical decomposition for noise suppression, hand-crafted features for texture characterization, statistical feature selection for dimensionality reduction, and a hybrid reinforcement learning classifier for the final decision. Each stage can be examined and improved independently, which is a significant advantage in medical applications where transparency and accountability matter.

The chimp optimization algorithm itself, introduced by Khishe and Mosavi in 2020, has gained attention for its balance of exploration and exploitation in the search space. By dividing the search population into four types of chimps—drivers, barriers, followers, and chasers—the algorithm navigates the parameter space efficiently, avoiding the premature convergence that plagues many metaheuristics. In the context of ChOAFQL, this means the fuzzy membership functions and Q-learning parameters are tuned with a level of precision that would be difficult to achieve through manual calibration or gradient-based methods alone.

The clinical implications of this work are considerable. An automated tool that can classify not only the presence but also the stage of glaucoma could serve as a screening aid in settings where specialist ophthalmologists are scarce, allowing general practitioners or technicians to flag patients who need urgent referral. Because the method relies on standard fundus photographs, which are inexpensive and widely available, it could be deployed in resource-limited environments without requiring specialized equipment such as optical coherence tomography scanners. The authors suggest that the proposed approach can serve as an effective and dependable tool for automated diagnosis and staging of glaucoma.

As with any machine learning system applied to medicine, careful validation on larger and more diverse patient populations will be essential before clinical deployment, and regulatory approval would require extensive testing. Nevertheless, the combination of near-99 percent accuracy, statistical significance testing, cross-validation, and robustness to image degradation makes this framework a compelling candidate for further development. The study, published in Volume 38 of Neural Computing and Applications as article number 819, adds to a growing body of evidence that hybrid computational intelligence—blending empirical decomposition, feature engineering, and nature-inspired optimization—can match or exceed the performance of purely deep learning approaches in medical image analysis, while offering greater interpretability at every step of the pipeline.

Subject of Research: Glaucoma identification

Article Title: Bi-dimensional empirical and wavelet decomposed hybrid chimp fuzzy Q-learning classifier for glaucoma identification

Article References: Pandey, G., Sharma, R., Kukker, A., & Negi, R. S. (2026). Bi-dimensional empirical and wavelet decomposed hybrid chimp fuzzy Q-learning classifier for glaucoma identification. Neural Computing and Applications, 38(20), Article 819. https://doi.org/10.1007/s00521-026-12489-y

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12489-y

Keywords: Bi-dimensional, empirical, wavelet, decomposed, hybrid, chimp, fuzzy, Q-learning, classifier, glaucoma, identification, scientific research