Deep learning models have demonstrated remarkable performance across the glaucoma's diagnosis and prognosis; however, deploying them in resource-constrained environments poses significant challenges. This research explores the balance between compression and accuracy preservation in specialist convolutional neural networks (CNNs) intended for CPU-based execution with minimal storage requirements. By employing pruning, knowledge distillation, quantization, and weight sharing, it is aimed to achieve maximal compression without compromising essential task performance. Resulting findings provide insights into the efficiency limits of model compression and its implications for real-world deployment. Additionally, the applicability of these compression techniques to Transformer-based architectures is examined throughout the work, which pose unique challenges due to their reliance on attention mechanisms. deep learning, glaucoma's diagnosis and prognosis, medical imaging optic nerve segmentation, fundus image analysis, model compression, pruning, knowledge distillation, quantization, weight sharing, transformer models, convolutional neural networks (CNN), low-resource deployment, edge computing, self-attention mechanisms, transfer learning, efficiency, optimization, neural network performance, sparse models, inference speed, binary classification. 1 Intelitsis'25: The 6th International Workshop on Intelligent Information Technologies & Systems of Information Security, April 04, 2025, Khmelnytskyi, Ukraine ∗ Corresponding author. † These authors contributed equally.
Glaucoma is the second most common cause of blindness and disability in the world [
The main difficulty in diagnosing glaucoma is that its symptoms usually appear only when vision has already deteriorated significantly. Due to the asymptomatic course and slow development of the disease, many people are unaware of the problem, which makes it difficult to detect and predict it early, especially in the initial stages. However, timely diagnosis can slow down the progression of the disease and prevent vision loss.
Since glaucoma causes retinal damage due to damage to the optic nerve head and increased
intraocular pressure, an important step in its detection is the segmentation of the optic disc. This
procedure is difficult due to the small size of the disc and possible blood supply disorders [
The standard diagnosis of glaucoma includes an assessment of the optic nerve head by examining
the retina, measuring intraocular pressure, analyzing visual fields, and taking into account other
related factors [
A promising area for improving the diagnostic process is modern information consulting and
diagnostic technologies. They help to improve medical decisions and the quality of treatment of
patients with retinal pathologies and are already used in numerous systems for diagnosing various
diseases [
Over the past decade, machine learning and deep learning methods have developed significantly,
not only improving the accuracy of glaucoma detection but also speeding up the processing of large
amounts of images and data, making the work of doctors easier [
Recent advancements in deep learning have led to increasingly large and complex models, making deployment on limited-resource environments such as edge devices and low-power CPUs impractical. While specialist models tailored for specific tasks offer high accuracy, they often retain redundant parameters that inflate model size and computational overhead. The challenge lies in optimizing these models for real-world applications while maintaining their effectiveness.
Compression techniques [
The balance between model size and performance has been widely studied. Han et al. [
Additionally, research by Bala and Eschbach [20] highlights the importance of optimized color spaces for improving neural network visibility in medical imaging. Their findings inform our preprocessing techniques for fundus image analysis.
Studies such as Wu et al. [18] emphasize the delicate balance between compression level and accuracy. Excessive pruning or quantization can introduce errors that degrade performance, highlighting the need for a strategic approach to compression. Knowledge distillation has also been explored as a means to retain essential knowledge in highly compressed models while mitigating accuracy loss.
Research has shown that specialized models tend to outperform general-purpose models in their respective domains [21-24]. This motivated selection of an image transformer model for optic nerve extraction, as transformers have demonstrated superior performance in medical imaging tasks. By leveraging their self-attention mechanisms, these models excel in capturing fine-grained spatial dependencies, which is crucial for precise segmentation of the optic nerve in fundus images.
CNN-based encoder-decoder architecture was utilized, structured as follows: 1. Initial feature extraction – VGG-16 backbone with four pooling layers. 2. Intermediate processing – custom layers for high-level feature abstraction. 3. Decoding – three decoder layers with skip connections to preserve spatial information.
This model is fine-tuned for binary classification of optic nerve localization in fundus images. Model is used for image segmentation into single-channel optic mask [25]. The model is tasked to learn high-level features, which, in practice, may further decompose into finer feature representations depending on the model's depth and granularity. Feature rough estimate is 17 discrete low-level features that the CNN must detect, process, and classify to correctly locate the optic nerve. Feature breakdown follows:
1. Brightness intensity contrast (1). 2. Circular shape detection (1). 3. Local texture gradient (1). 4. Edge detection (1). 5. Color consistency (1).
Feature 2: Conjunction of several blood vessels: 6. Vessel edge detection (1). 7. Vessel thickness estimation (1). 8. Vessel directionality (1). 9. Vessel junction detection (1). 10. Vessel curvature (1).
Feature 3: Rounded dark border around the bright spot: 11. Contrast difference (1). 12. Border shape (1). 13. Edge sharpness (1). 14. Shadowing effects (1).
Feature 4: Not on the black part of the image: 15. Background segmentation (1). 16. Foreground detection (1). 17. Image region classification (1).
Image describing structure of base-64 transformer model is provided in Figure 1.
To enhance image quality and improve feature extraction, histogram equalization and CLAHE (Contrast Limited Adaptive Histogram Equalization) in the LAB color space are applied. These techniques help normalize contrast and highlight relevant structures in the fundus images. The choice of the LAB color space is inspired by research from [26, 27], which demonstrates that optimized color spaces can improve neural network visibility and feature extraction in medical imaging.
Iterative structured pruning is being applied to model, progressively removing low-importance weights while monitoring accuracy loss, resulting model was fine-tuned. In this article pruning was pretty aggressive (50%). Dropout layers (30% rate) were also applied to force model to compress knowledge and decrease amount of knowledge lost on pruning application, lessening retrain time and shortening preparation for further compression.
A smaller student model is trained to mimic the larger teacher model, learning its essential features while significantly reducing parameter count. In this article larger first iteration model was used as teacher model for all further iterations of student model learning.
Post-training quantization was employed to convert 32-bit floating-point weights to 8-bit integers, reducing memory consumption. Spoiler: memory consumption remained unchanged, inference time remained largely unchanged (5% speed increase), space on disk actually increased (why?).
K-Means clustering was used to group similar weights, replacing them with shared values to reduce parameter storage. Spoiler: ineffective for accuracy, inference time unchanged, memory consumption unchanged.
Transfer learning was employed to teach the first large iterations of the transformer faster, leveraging pre-trained models to accelerate convergence and improve performance. Prior research has demonstrated the effectiveness of transfer learning in deep learning applications, particularly in medical imaging tasks, where pre-trained architectures significantly enhance detection accuracy and efficiency [28].
In addition to the CNN-based approach detailed above, current research evaluates an image transformer model for optic nerve extraction. Transformer architectures are renowned for their selfattention mechanisms, which excel at capturing long-range dependencies and fine-grained contextual relationships-a quality that is particularly beneficial for specialized tasks in medical imaging. In the context of optic nerve segmentation, transformers are capable of discerning subtle variations in image features that are crucial for accurate detection.
While recent literature has introduced a range of advanced compression strategies tailored to transformers (example - low-rank factorization [29, 30] and sparse attention mechanisms [31, 32]), current work does not implement these methods. Instead, the transformer model is utilized in its standard form to serve as a benchmark for specialized model performance. This approach allows to directly compare the efficiency gains achieved through CNN compression techniques - pruning, knowledge distillation, quantization, and weight sharing - with the inherent advantages of transformer-based models operating under similar CPU constraints.
By focusing on this comparison, we seek to determine where the transformer’s compressed performance can match accuracy of base transformer model. This evaluation is critical to overall theme of liminal efficiency, where we explore the threshold at which aggressive compression techniques can be applied without incurring unacceptable losses in accuracy, while considering the trade-offs between generalist and specialist model architectures. Liminal efficiency is defined as <5% accuracy loss from starting model without losing binary feature classification metrics and increasing performance.
For experiments we use datasets ODIR-5K [33] – local ophthalmologist-provided datasets, which consists of 500 random images selected for training and evaluation.
We calculated following metrics: inference time, total training time, number of training epochs for sufficient results (here training epoch is one run over training image set), evaluation prediction count, total true positive predictions, total true negative predictions, total false positive predictions, total false negative predictions, additional false detections, model size, and complex metrics: accuracy, recall, and F1-score.
For testing the results 871 images tested in a CPU environment.
Experiments are conducted on model bases of 64, 32, 16, 8. Iteration model graphic images for reference of differences in iteration for base 32 and base 16 are displayed on Figure 2 and Figure 3 respectively (difference is in thickness of convolution blocks – these are multipliers of base).
Training was conducted using k-fold cross-validation algorithm with resulting loss being extracted from evaluation fold. Additionally, to preserve edge cases knowledge a set of highest loss cases in maximum size of 1 fold was assembled and used every step to additionally randomly select a batch to train upon using weighted random where higher loss cases are selected more often. Both knowledge distillation and pruning were trained on similar edge cases.
Pruning experiment results are provided in Table 1. Pruning experiments involve following: 1. Record metrics for each iteration, epoch and training session. 2. At iteration 1 transfer-learning is applied to model encoder blocks 1 through 4, then dropout is applied at each activation except the very last block using dropout-30% layers. After that model is trained to find optic nerve until sufficient results are obtained. 3. At subsequent iterations: previous iteration model is pruned by 50% with dropout turned off, dropout is turned back on after 1 training epoch and model is subsequently fine-tuned until sufficient accuracy is obtained. 4. Repeat for following “bases” of transformer: 64 (initial), 32, 16, 8. We stopped at base 8 because model stopped converging fast enough requiring obnoxious amount of training time (recorded results are results of first converged training, however last iteration did not converge at all).
Knowledge distillation experiment results are provided in Table 2. Knowledge distillation experiments involve following: 1. Record metrics for each iteration, epoch and training session. 2. At iteration 1 transfer-learning is applied to model encoder blocks 1 through 4. After that model is trained to find optic nerve until sufficient results are obtained. 3. At subsequent iterations: use results from iteration 1 as training target. 4. Repeat for following iteration “bases” of transformer: 64(initial, step 2), 32, 16, 8. I stopped at base 8 because model(again) stopped converging fast enough, requiring obnoxious amount of training time. This time base 8 model was trained for another 40 epochs till sufficient accuracy.
Resulting images that denote edge cases handling are displayed in Figures 4-8.
Weight sharing testing involve following: Use final model from training for certain model base and apply weight sharing, fine-tune and reapply weight sharing till accuracy is acceptable.
Results: unusable, breaks detection. Training and weight sharing after each epoch did not improve result. Applying weight sharing after every training did not improve resulting accuracy. Applying weight sharing after several epochs passed did not improve accuracy. Issue with weight sharing was that there were too many interconnected target clusters for good enough selection of optic nerve. Issue example provided in Figure 9.
Analysis of experiments’ results shows that the x16 base dropout pruning variant is a liminal point in terms of accuracy/performance ratio. It exhibits negligible accuracy loss compared to previous results while being smaller in size and 5.8 times faster than the x64 base transformer. Moreover, it retains edge case detection capability, indicating superior knowledge retention and fine-tuning potential. This variant remains undertrained and can be fine-tuned further along with unconstraining with setting dropout to 0. Alternatively, data from knowledge distillation experiments suggest that the x8 base transformer has performed well in general cases. Therefore, an x8 nodropout pruned variant may be obtained from the x16 dropout 30 variant by pruning it to x8 base and fine-tuning it with dropout set to 0 for maximum performance and knowledge retention.
During the preparation of this work, the authors used X-GPT-4o in order to: Grammar and spelling check. After using this service, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.