Much research in glaucoma detection, most has relied on discrete CNNs and simplified clustering techniques

The gap

Evidence profile

Sourced from the future work and discussion and conclusions and limitations of the source papers, classified as general, spanning 7 journals. Those papers have been cited 2 times in total.

Supporting evidence — 8 representative gaps

• Explainable neuro-symbolic artificial intelligence for automated interpretation of corneal topography and early keratoconus detection (2026) · Frontiers in Artificial Intelligence · cited 1× · doi

  Future research should focus on validating the proposed frame- work using larger multicenter datasets that include diverse populations and imaging platforms. Integration of additional diagnostic modali- ties, such as corneal tomography, biomechanical measurements, or anterior segment optical coherence tomography, may further enhance the robustness of early keratoconus detection. In addition, the neuro- symbolic reasoning architecture may be extended to other ophthalmic diagnostic tasks, including glaucoma risk assessment, corneal dystro- phy classification, and refractive surgery planning. From a broader perspective, combining symbolic clinical knowl- edge, deep learning representations, and explainable language models represents a promising direction for the development of transparent and trustworthy medical AI systems. Such hybrid approaches may help address one of the central challenges in clinical AI: balancing predictive performance with interpretability and clinical accountability.

  generalfuture work
  Keywords: clinical diagnostic corneal tomography symbolic future focus validating proposed frame using larger multicenter datasets include
• Explainable AI-driven diagnosis model for early glaucoma detection using grey-wolf optimized extreme learning machine approach (2026) · PLOS Computational Biology · doi

  Our proposed manuscript presents an improved explainability framework called GlaucoXAI (Glaucoma Explainable Artificial Intelligence). GlaucoXAI aims to enhance our understanding of deep learning networks’ behavior by utilizing advanced visualization techniques, such as attention maps. As a post hoc tool, GlaucoXAI can be applied to any existing deep neural models, offering significant insights into their operations. Our two case studies highlight the importance of integrating explainable AI (XAI) techniques in medical image analysis. Additionally, GlaucoXAI facilitates the Extreme Learning Machine (ELM) classifier for glaucoma detection. Our findings underscore the crucial role of XAI in medical imaging tasks, helping to improve the comprehensibility of machine learning models and accelerating their adoption by medical professionals. This paper introduces a sophisticated computer-aided design (CAD) model specifically designed for categorizing glaucoma and healthy images. The model effectively identifies relevant features in fundus images by utilizing a fast discrete curvelet transform with a wrapping (FDCT-WRP) process. To enhance feature reduction, we apply a combination of Principal Compo- nent Analysis (PCA) and Linear Discriminant Analysis (LDA), resulting in a set of reduced and more prominent features. Subsequently, the CAD model employs the IMGWO-ELM, a faster learning algorithm, to train the Single-Layer Feed- forward Network (SLFN). We rigorously evaluate the CAD model’s classification performance across two standard fundus image datasets. The experimental results demonstrate that the proposed CAD model achieves superior classification performance with fewer features compared to existing models. PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1013139 May 04, 2026 30 / 34 In future research, we plan to test the efficacy of our approach for generalization across various imaging modalities. Another potential avenue of exploration is the hybridization of ELM with a less parameter-based optimization algorithm, assessing its effectiveness in multi-class classification tasks. Additionally, the paper suggests considering deep learning algorithms as viable alternatives to the proposed model. Our future research will also focus on quantitatively evaluating XAI methods. This evaluation aims to assess the effectiveness of sensitivity maps generated by these methods and their correlation with deep learning accuracy metrics. We plan to conduct further experiments in the realm of multi-modal glaucoma detection to enhance our understanding. Moreover, we aim to investigate the potential for extracting quantitative features, such as tumor volume and centroid, from these explanation methods.

  generalfuture work
  Keywords: learning model glaucoxai glaucoma deep features proposed enhance models medical classification explainable aims understanding utilizing
• Explainable AI-driven diagnosis model for early glaucoma detection using grey-wolf optimized extreme learning machine approach (2026) · PLOS Computational Biology · doi

  Future work will focus on extending the model for multi-class and multi-modal glaucoma classification, integrating additional clinical data, and exploring more adaptive optimization strategies. Variability in the explanations produced by different Explainable AI (XAI) techniques also presents challenges, as consistency in interpret- ability remains an open issue.

  generaldiscussion
  Keywords: multi future focus extending model class modal glaucoma classification integrating additional clinical exploring adaptive optimization
• Development and Comparative Evaluation of Hybrid EfficientNet-CapsNet Architecture for Glaucoma Diagnosis in Resource-Limited Settings (2026) · Journal of Al-Qadisiyah for Computer Science and Mathematics · doi

  While there has been much research in glaucoma detection, most has relied on discrete CNNs and simplified clustering techniques, leading to inconsistent results due to their limited generalizability across different imaging modes and their inability to preserve the anatomical state of the optic nerve head.

  generalconclusions
  Keywords: there glaucoma detection relied discrete cnns simplified clustering techniques leading inconsistent limited generalizability across different
• Artificial intelligence for early detection of diabetic retinopathy: A vision transformer-based approach (2026) · PLOS One · cited 1× · doi

  6.1 Summary of key findings The key findings in this study can be presented as follow: • A model for DR identification called Compact Convolutional Transformer for Automated Diagnosis (CCTAD) is proposed in this study. • An improvement of feature extraction, by integrating convolutional tokenization with self-attention mechanisms, that effectively capturing both local and global image features. • According to obtained results, we can ensure that the proposed CCTAD have achieved high precision, recall, and F1-scores across different severity levels, outperforming conventional CNN-based models in terms of classification accuracy and efficiency. • Additionally, preprocessing techniques such as contrast enhancement and noise reduction significantly contributed and greatly helped to enhance model performance. 6.2 Future research directions To further improve CC TAD’s functionality and applicability, future studies will concentrate on interesting directions as opti- mizing computational efficiency, integrating the system into real-world deployment in healthcare applications. • Model Optimization:Various techniques, like Pruning, quantization, and knowledge distillation, can be used as methods of reducing model complexity and enabling deployment on resource-constrained devices. • Clinical Validation: The generalizability and reliability of CCTAD in real healthcare settings will be evaluated through extensive validation on varied clinical datasets and real-world patient data. • Real-world feature: In real-world screening programs, a significant portion of fundus images may be ungradable due to blur, low illumination, or motion artifacts. Although the datasets used in this study predominantly contained high-quality, expert-verified images, future implementations of the proposed CCTAD system will integrate an auto- mated image-quality assessment module. This component will use sharpness metrics (e.g., variance of Laplacian), contrast evaluation, and deep-learning–based IQA networks to identify and exclude ungradable images before anal- ysis. Incorporating this module will enhance the reliability and safety of automated DR screening in large-scale or teleophthalmology settings. • Deployment in Healthcare Settings:Ensure that AI systems and medical professionals can work together seamlessly by investigating the integration of CCTAD into automated screening systems and telemedicine platforms.

  generalfuture work
  Keywords: cctad real model world automated proposed future deployment healthcare settings screening images convolutional feature integrating
• A Review of Deep Learning in Medical Image Recognition for Glaucoma (2026) · International Journal of Public Health and Medical Research · doi

  Deep learning techniques have demonstrated significant efficacy in the recognition of medical images for glaucoma. Automated detection models based on convolutional neural networks (CNNs) can successfully perform optic disc and cup segmentation as well as glaucoma classification, achieving performance levels that meet or even surpass those of professional ophthalmologists. Extensive research has been conducted in this field, yielding various network architectures, effective feature extraction methods, and advanced model improvement strategies. However, several challenges remain, including the scarcity of sufficiently large datasets, limited model generalization capabilities, and a lack of interpretability. Future research should prioritize the following areas. First, large-scale, multi-center, and standardized glaucoma image databases must be established to facilitate extensive training and testing. Second, lightweight network architectures and model compression techniques should be developed to reduce computational overhead, thereby enabling the deployment of models on mobile or edge computing devices. Third, model interpretability must be improved by designing visualization tools that demonstrate how decisions are made, which will foster greater clinical trust in AI-assisted diagnoses. 79 Fourth, multi-source information fusion and cross-modal learning should be investigated to fully leverage complementary data from various examination modalities. Finally, privacy-preserving technologies, such as federated learning, should be utilized to enable collaborative modeling across multiple hospitals while strictly protecting patient privacy. With continued technological advancements, deep learning is expected to play an increasingly vital role in the early detection, accurate diagnosis, and personalized treatment of glaucoma. Ultimately, these developments will effectively mitigate the incidence of glaucoma-induced blindness and safeguard public visual health. REFERENCES [1] Shyamalee, T., & Meedeniya, D. (2022). Glaucoma detection with retinal fundus images using segmentation and classification. Machine Intelligence Research, 19(6), 82-99. https://doi.org/10.1007/s11633-022-1354-zMachine In... [2] Devi, M. C., & Ramaswami, M. (2025). Efficient hybrid deep learning network model for segmentation and classification of heart angiographic images. Nano Biomedicine and Engineering, (3), 114 -129. [3] Hui, B., Liu, Y., Qiu, J., Cao, L., Ji, L., & He, Z. (2021). Study of texture segmentation and classification for grading small hepatocellular carcinoma based on CT images. Tsinghua Science and Technology, (2), 67 -75. [4] Salama, W. M., & Aly, M. H. (2022). Framework for COVID-19 segmentation and classification based on deep learning of computed tomography lung images. Journal of Electronic Science

  generalfuture work
  Keywords: learning glaucoma images segmentation classification model deep detection based network techniques models extensive various architectures
• Artificial Intelligence in Opthamology:A study on different AIML approaches for Glaucoma prediction (2026) · Zenodo (CERN European Organization for Nuclear Research) · doi

  challenges. There are concerns regarding bias, reproducibility, and equity because most of the studies rely on datasets which are small or not diverse enough. The unequal distribution of disease prevalence and the underrepresentation of specific ethnic groups intensify these issues. To address these challenges, it is essential to develop larger and more varied datasets, preferably utilizing collaborative methods like Federated Learning to uphold patient privacy. Furthermore, specific techniques like fair identity normalization are being developed to ensure equitable screening performance across all populations. excel Ultimately, each method has its own advantages and disadvantages. at Imaging-focused models assessing structure but often miss functional elements. EHR-driven systems facilitate large-scale prescreening but lack ophthalmic specificity. Sensor-based approaches offer real-time physiological insights but face limitations due to device availability and cost. This underscores the necessity for multimodal integration, where structural, functional, and systemic data are merged to better represent actual clinical practice. Such approaches hold the greatest promise for enhancing diagnostic precision, predicting disease progression, and fostering the clinical acceptance of AI-driven glaucoma systems.

  generallimitationsevidence 5/5
  Keywords: challenges datasets disease specific like functional driven systems approaches clinical there concerns regarding bias reproducibility
• A Robust MRI-Based Brain Tumor Diagnosis Framework Integrating U-Net, EfficientNet-B7, and Grad-CAM Visualization (2026) · International Journal of Drug Delivery Technology · doi

  regarding for the best choices automatic discovery of features and end-to-end learning process. Among deep learning methods, U-Net became one of the most efficient tools for biomedical image segmentation due to the implementation of an encoder-decoder architecture and skip connections that allow retaining spatial features. On the other hand, image classification models such as EfficientNet family of architectures perform better by employing compound scaling strategies and simultaneously optimizing network depth, width, and resolution. In this regard, EfficientNet-B7 is considered one of image classification tasks. However, some recent progress has been made in the context of transformer architectures and Vision Mamba, allowing longer dependencies within medical images. Despite the advantages gained by the new technologies, transformer models still require huge computational resources and large training datasets. Also, many existing methods focus only on one it segmentation or task, be classification, but not both. Another important challenge that exists when designing medical AI systems is that of interpretability. Deep learning models can often be viewed as "black boxes," meaning that physicians find it difficult to determine how their predictions. In particular, Explainable Artificial Intelligence (XAI), including Grad-CAM, provides visualized interpretations through showing parts of input data that help form decisions.

  generallimitationsevidence 5/5
  Keywords: learning image features classification models efficientnet deep segmentation architecture architectures transformer medical grad regarding best

Questions about this gap