Convoluted Neural Networks, CNN, AI Deep Learning Models and their Role in the Diagnosis of Glaucoma

Glaucoma Epidemiology

The global prevalence of glaucoma in 2020 was approximately 79.6 million individuals affected with open angle and closed angle glaucoma [1]. About 4.5 million individuals worldwide are blinded by glaucoma, such that glaucoma worldwide is the second most common cause of loss of vision [2].

Glaucoma Diagnostics Plus CNN

While intraocular pressure, visual field testing, (especially with the digital field analyzers), and Optical Coherence Tomography, OCT, and gonioscopy are all valuable tools in the diagnosis of glaucoma, the addition of CNN to analyze fundus photo and OCT images of the optic nerve is a valuable, relatively inexpensive, noninvasive, adjunctive technique to gain earlier diagnostic knowledge of individual, familial and population at risk levels in this silent but threatening disease.

The glaucoma diagnostic leverage with the addition of CNN has widespread implications in terms of early: personal care adjustments, community access in underdeveloped areas, and management of lack of sustained glaucoma medication adherence over extended periods.

The CNN optic nerve layers from fundus or OCT images, with their guided deep learning and, forward and backward numeric tensor propagations, can iteratively decipher glaucoma patterns, at close to experienced ophthalmologist level.

CNNs Deep Historical Elements

CNNs historically go back to Hubel and Wiesel who discovered that the visual cortex has “simple cells” that respond to bars and edges and “complex cells” (less position sensitive) responding to bars/slits [3]. Torsten Wiesel’s 1981 Nobel Lecture begins with “In the early sixties, having begun to describe the physiology of cells in the adult cat visual cortex, David Hubel and I decided to investigate how the highly specific response properties of cortical cells emerged during postnatal development” [4]. “This modest statement belies the tidal wave of experiments on developmental brain plasticity that was initiated by their publications in 1962, continues to this day, and extends well beyond the occipital lobe to virtually all sensory areas, motor areas, and to “higher centers” involved in learning, memory, and decision making” [5]. This neurophysiological tapestry of simple to more complex sensory responses to visual stimuli morphing into contextually plastic and developmental iterations resulting in motor “learning, memory and decision making” forms the basis of the granular elements of CNN.

Table 1:Some Operational Roles of CNN Layers/Terms in Diagnosing Glaucoma.

CNNs Technical Evolution, Historical Elements

In 1979 and 1980 K Fukushima published the basic CNN architecture of convolutional layers and downsampling layers. (Fukushima first introduced non-linear activation units ReLU, rectified linear units to learn edges and complex patterns [7,6]) in 1969). His network was trained by unsupervised learning rules and was 100x more expensive than in 1989 and a billion times more expensive than today [6,7]. Fukushima made a video in 1986 of a CNN that recognizes handwritten digits [7]. In 1987, Alex Waibel (a German researcher working in Japan) trained supervised learning, weight sharing neural networks with 1-dimensional convolutions by Linnainmaa’s 1970 backpropagaton algorithm to recognize speech. A similar proposal by Homma et al. introduced the “convolution” terminology to neural networks. Convolving means taking a small numerical sample of a pattern in an image [7].

Both Wei Zhang and Yann LeCun, et al. (at Bell Labs) independently developed backpropagation-trained CNNs, (Zhang for Chinese letter character recognition and LeCun for zip codes) [7]. Backpropagation generates gradients or derivatives backward from the output layer to correct how much each weight and bias needs to be reduced to minimize the forward propagation overages; thus, using differential calculus to find optimal layers adjustments to fine tune the forward propagation of weights and biases, the back propagation allows for weights to be updated and reduce prediction errors and avoid redundant calculations [7,8].

Table 2: CNN lexicon of terms.

The Centrality of the optic nerve changes to glaucoma diagnosis; non-AI glaucoma diagnostic accuracy, and false positives that can look like glaucomatous optic atrophy

CNNs, after importing an optic nerve image on a computer, for example, can convert it to a numeric array and then with coded directions through guided neural net layers iteratively identify glaucoma patterns in the optic nerve. These glaucoma associated patterns are learned so well from the CNN models, that their accuracy rivals that of experienced ophthalmologists.

In 2002 a randomized controlled study in ocular hypertension found: 1. Ocular hypertension is a major risk factor in developing primary open angle glaucoma. 2. Treatment of ocular hypertension with pressure lowering medication delayed or prevented glaucoma onset. 3. Over half the patients who developed glaucoma showed optic nerve changes, without early visual field loss, emphasizing the need for optic nerve monitoring alongside ocular pressure checks. 4. Lowering intraocular pressure helps protect the optic nerve and slows down the damage rate. 5. Thinner central corneal thickness is another significant risk factor for glaucoma [17].

Therefore, by using CNN to detect glaucoma optic nerve changes, even before visual field loss, early diagnosis, treatment and patient counseling can be achieved. The latter, of course, contextually demands a critical evaluation of the individual patient: the patient’s medical history, physical data, lab data, and clinical and familial background.

That said, over the last decade several papers using CNN on optic nerve images found glaucoma patterns, and their level of identification compared well to that of trained ophthalmologists.

However, to start with, let’s examine a non-AI baseline accuracy level in trained eyecare settings, in a “masked performance study” in Scotland. Patients suspected of having glaucoma, within one month, underwent a full ophthalmic assessment in both a newly established community led glaucoma management location, AGO, and a consultant led eye hospital. Agreement between the AGO and the consultant unit in diagnosing glaucoma was 89%. Agreement between trainee ophalmologists and the consultant unit was 83%. The accuracy of optometrists in detecting glaucoma in this population was good with a specificity of .93, but lower for sensitivity at .76. There was no difference in sensitivity between AGO and junior ophthalmologist [18]. The diagnosis of glaucoma in this study rested on optic nerve changes and or visual field abnormalities [18].

The other side of the coin, false positives, may be neuroophthalmological conditions that mimic glaucoma and result in misdiagnosis [19]. These include: ischemic optic neuropathy, 25%, compressive optic neuropathy, 18.7%, hereditary optic neuropathy, 18.7%, and congenital optic neuropathy, 2%, from a study of 68 patients enrolled with neuroophthalmological diseases screened from a single Eye Clinic within a 24-month period.

The researchers selected the eyes with pre-defined glaucoma criteria:

Vertical cup-to-disc ratio greater than or equal to 0.6, asymmetric disc ratio greater than 0.2 between the two eyes, presence of localized retinal nerve fiber layer and or neuroretinal rim defects and disc hemorrhages.

The images were mixed randomly and a masked glaucoma expert was asked to distinguish if each patient exam derived from a patient with glaucoma or a neurophthalmological condition. “Based on the analysis of fundus photographs and HVF, (Humphrey Visual Field) tests, 25% of these were misdiagnosed as glaucoma (two ischemic optic neuropathies and two congenital optic disc anomalies). Conversely, 11.9% of the glaucomatous neuropathies were misdiagnosed as neurophthalmological disorders. Overall, the glaucoma specialist correctly diagnosed 84.5% of the eyes” [19].

References 18 and 19 illustrate the need for extensive analysis and for thorough individual patient data review for a glaucoma diagnosis to be made, even in the setting of prima facie, conventional clinical markers of glaucoma.

With that introduction, we can explore the results of CNN glaucoma pattern recognition in about a decade worth of papers, using different CNN deep learning models to detect glaucoma and their relative accuracies.

CNN Layers, Activators, features to develop a Model to test, validate and train optic nerve images for normal/ glaucoma prediction, Procedural Considerations

After downloading a series of “procedural software packages”, the initial activity of the CNN neural network coder is to import the image or data set of images. In this setting the image(s) will be either fundus photo(s) or an OCT, Optical Coherent Tomogram, image(s) of an optic nerve with an identifying label.

The programmer can then elect to use an established, well vetted deep learning platform with its parameters or implement a custom platform with additional parameters. In any case, the platform will serve as a base for the CNNs network to import the image data set, convolve it, add weights, adjust it, and optimize it, prior to train, validate, test and predict dynamics. Yet additional modulatory features that the coder wishes to adjust /add called hyperparameters can also be implemented. How different the CNN model is from the actual desired outcome or its loss function as well as and the training model’s accuracy will determine just how much adjustment the training model requires.

One or more convolutional neural network layers, each with a specific function, such as feature extraction, segmentation, classification [12, 20], will without the knowledge or assistance of the coder/observer process large amounts of visual data. The CNN layer will capture an extracted kernel of the image (which is a smaller sample of numerical matrix feature) for the character of the item it is detecting and connect it to others of a similar nature. These are joined and either made to relate, or not, to other layers. A hierarchy of layer connections results in optimizing the input and creating an architecture of learning matrices with corresponding numeric outputs called logits that represent complex digital outputs of salient patterns in the image. These logits or quantized outputs can then be used for training, testing, and plotting the image data, and evaluating the data for inference and predictions [12, 20,21,22].

To reduce the large spatial dimensions of the convolutional layers, pooling layers downsize the feature data maps but still preserve the essential data [12,22].

To introduce non-linearity into the CNN model, the ReLU, rectified linear unit, is used; other activation functions are sigmoid and tanh; sigmoid maps from 0 to 1 and tanh maps from -1. to 1; tanh is zero centered which leads to faster image data convergence and better hidden layer performance. Sigmoid, on the other hand, is better for working the output layer in binary classification where you need a probability of 0 to 1 [21].

Connected layers join the learned features of the other layers, combine spatial information and transform it into a format suitable for classification and provide the basis for making relationships and enabling high level decisions [12, 22].

Thus, the CNN architecture provides a learning platform for image pattern recognition, representation, segmentation, classification and automatic aggregation to facilitate identification of normal vs glaucoma, e.g. cupping, neuronal rim thinning, and retinal nerve fiber layer features. This collection of CNN aggregated images can be used to develop, test, and train data sets of new images that can (with a background of good accuracy validation) be used to predict normal from abnormal optic nerves in successive glaucoma test samples [9,11,12,22].

Table 3 illustrates some of the papers with greater than or equal to 90% accuracy in diagnosing glaucoma with various CNN methods. These papers reflect glaucoma detection from fundus photos (from conventional fundus photo units as well as a portable ophthalmoscope combined with a smart phone) as well as glaucoma detection from 3D OCT images of the optic nerve, and recurrent neural network images of sequential video fundus images.

Table 3:Papers with >90% Accuracy Diagnosing Glaucoma with CNN.

Exploring the realm of refining CNN choices from the established CNN glaucoma deep learning models requires experienced investigation in not only the basic parameters of the established model, but also the adaptation of hyperparameters by the programmer for the new model. This type of experienced investigation, (at least until agentic modelling in this area can be validated) is definitely one of the drawbacks of the current CNN application to the diagnosis of glaucoma in a given data set, especially if the data set needs adaptation to a particular CNN model that may not have been used with that data set previously.

Collection and transferring images, as was evident in Bragance, et al. (using e.g. portable ophthalmoscope and smart phone) can also be a challenge. However, if functional code programs or agentic programs connect image: intake, preparation for CNN, integration with CNN and development of optimal outputs for glaucoma predictions, then even in an eye clinic setting portable ophthalmoscopes and smart phones may be used in the future.

That said, agentic deep learning is progressing at a nice pace. With the laudable ideal of streamlining CNNs and their linkage to diagnosing glaucomatous optic nerve photos for millions in need, the technology may be short in coming.

From the above work on using CNNs to detect glaucoma from fundus or OCT images of the optic nerves, it would appear that it is only a matter of time for the conventional eye clinic to integrate CNN with tonometry and visual fields for glaucoma detection. As fundus or OCT imaging gains greater bandwidth with CNN integration, in house earlier glaucoma detection can become a reality even in more rural areas.

None.

Monica chatbot, Microsoft Copilot, and Google Gemini were used to cross check CNN methodology definitions with references used; also, these sources were helpful in understanding Accuracy metrics in different papers.

1. https://glaucoma.ph
2. Y C Tham, X Li, TY Wong, HA Quigley, T Aung, et al. (2014) Global prevalence of glaucoma and projections of glaucoma burden through 2040: a systematic review and meta-analysis. Ophthalmology 121(11): 2081-2090.
3. Hubel, DH, Wiesel, TH (1962) Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex. J Physiol 160(1): 106-154.
4. http://nobelprize.org/nobel_prizes/medicine/laureates/1981/wiesel-lecture.html
5. Martha Constantine Paton (2008) Pioneers of cortical plasticity: six classic papers by Wiesel and Hubel. Journal of Neurophysiology 99: 6.
6. Fukushima K (1980) Neocognitron: A self-organizing neural network model for mechanism of pattern recognition unaffected by shift in position. Biological Cybernetics 36(4): 193-202.
7. Jürgen Schmidhuber (2025) Who invented convolutional neural networks.
8. Y LeCun, B Boser, JS Denker, D Henderson, RE Howard (1989) Backpropagation applied to handwritten Zip Code recognition. Neural Computation 1(4): 541-551.
9. Yamashita R, Nishio M, Do RKG, Togashi K (2018) Convolutional neural networks: an overview and application in radiology Insights Imaging 9(4): 611-629.
10. Sarvamangala DR, KuLkarni RV (2022) Convolutional neural networks in medical image understanding: Evol Intell 15(1): 1-22.
11. Raghavendra U, Fujita H, Bhandary SV, Gudigar A, Tan JH (2018) Acharya UR. Deep convolution neural network for accurate

diagnosis of glaucoma using digital fundus images, Inf Sci (Ny) 441: 41-49.

12. Haja SA, Mahadevappa V, Romanian (2023) J Ophthalmol 67:(3): 222-237.
13. Chen J, Li S, Bai Q, Yang J, Jiang S, Miao Y (2021) Classification Algorithms based in Convolutional Neural Networks, Remote Sens (Basel) 22: 4712.
14. Ananda, Karabag C, Ter Sarkisov A, Alonso E, Reyes Aldasoro CC (2020) Radiography classification: A comparison between eleven convolutional neural networks. In: 2020 Fourth International Conference on Multimedia Computing, Networking and Applications (MCNA). IEEE: 119-125.
15. Alzubaidi L, Zhang J, Humaidi AJ, (2021) Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. J Big Data 8(1): 53.
16. Camara J, Neto A, Pires IM, Villasana MV, Zdravevski E, et al. (2022) A comprehensive review of methods and equipment for aiding automatic glaucoma tracking. Diagnostics (Basel) 2(4): 935.
17. Kass MA, Heuer DK, Higgenbottom EJ, Johnson CA, Keltner JL (2002) The Ocular Hypertension Treatment Study: a randomized trial determines that topical ocular hypotensive medication delays or prevents the onset of primary open-angle glaucoma, Arch Ophthalmol 120(6): 701-713.
18. Azuara Blanco A, Burr J, Thomas R, Maclennan G, McPherson (2007) The accuracy of accredited glaucoma optometrists in the diagnosis and treatment recommendation for glaucoma, British J Ophthalmol 91(12): 1639-1643.
19. Dias DT, Ushida M, Battisella R, Dorairaj S, Prata ES, (2017) Neurophthalmological conditions mimicking glaucomatous optic neuropathy: analysis of the most common causes of misdiagnosis, BMC Ophthalmol.
20. Ribeiro E, Uhl A, Wimmer G, Häfner M, Exploring Deep Learning and transfer learning for colonic polyp classification. Comput Math Methods Med: 6584725.
21. Bourke D, learnpytorch.io, Learn Pytorch in a day, youtube.com
22. Sarvamangala DR, Kulkarni RV (2022) Convolutional neural networks in medical image understanding: a survey. Evol Intell 15(1): 1-22.
23. Shoukat A, Shahzad A, Syed AH, Hassan SA, Igbal S (2023) Automatic Diagnosis of glaucoma from retinal images using deep learning approach, Diagnostics (Basel) 13(10): 1738.
24. Akbar S, Hassan SA, Shoukat A, Alyami J, Bahaj SA (2022) Detection of microscopic glaucoma through fundus images using deep transfer learning approach, Microsc Res Tech 85(6): 2259-2276.
25. Bragance CP, Torres JM, Soares CPDA, Macedo LO (2022) Detection of Glaucoma on Fundus Images Using Deep Learning on a New Image Set Obtained with a Smartphone and Handheld Ophthalmoscope, Healthcare (Basel) 10(12): 2345.
26. Maetschke S, Anthony B, Ishikawa H, Wollstein G, Schuman J (2019) A feature agnostic approach for glaucoma detection in OCT volumes, PLoS One 14(7): e0219126.
27. George Y, Anthony BJ, Ishikawa H, Wollstein G, Schuman JS, (2020), Attention-guided 3D-CNN Framework for Glaucoma Detection and Structural-Functional Association using Volumetric Images, IEEE J Biomed Health Inform 24(12): 3421-3430.
28. Gheisari S, Shariflou S, Phu J, Kennedy P, Agar A (2021) Sci Rep 11: 1.