⚡ electrical engineering

Retinal Malady Classification using AI: A novel ViT-SVM combination architecture

This paper proposes a novel hybrid architecture combining Vision Transformers and Support Vector Machines to automate the early detection and classification of Macular Holes, Central Serous Retinopathy, and Diabetic Retinopathy using Optical Coherence Tomography scans.

Original authors: Shashwat Jha, Vishvaditya Luhach, Raju Poddar

Published 2026-04-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shashwat Jha, Vishvaditya Luhach, Raju Poddar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your eye is like a high-definition camera, and the retina is the film inside it. Sometimes, that film gets damaged, leading to blurry vision or even blindness. The three main "bugs" in the system this paper tackles are:

Diabetic Retinopathy: Like rust forming on the tiny pipes (blood vessels) inside the camera because of too much sugar in the blood.
Macular Holes: A literal tear or hole in the center of the film where the image is clearest.
Central Serous Retinopathy: A bubble of fluid that lifts the film away from the camera body, distorting the picture.

Doctors usually look at these scans (called OCT scans) with their own eyes to diagnose the problem. But human eyes get tired, and there are too many scans to check manually. This paper asks: Can we build a robot brain to do this faster and better?

The Old Way vs. The New Way

To answer this, the researchers tried three different "robot brains" (AI models) to sort the scans into four categories: Normal, Diabetic Retinopathy, Macular Hole, and Central Serous Retinopathy.

1. The Old Reliable: VGG-16

Think of this as a classic, hardworking detective who has been around for a long time. It looks at an image pixel by pixel, like reading a book one letter at a time. It's reliable but a bit slow and misses the big picture.

Result: It got about 83% of the cases right. Good, but not great.

2. The Modern Star: Vision Transformer (ViT)

This is the super-observant artist. Instead of reading letter by letter, it looks at the whole painting at once. It uses a technique called "Attention," which means it knows exactly which part of the image is important (like the hole in the retina) and ignores the background noise.

Result: It got 88% right. Much better than the old detective!

3. The Winning Team: ViT-SVM (The Hybrid)

This is the paper's big idea. They took the Modern Star (ViT) and paired it with a Mathematical Calculator (SVM).

Here is the analogy:

The ViT (The Artist): Looks at the blurry scan and says, "Hey, I see a weird pattern here that looks like a Macular Hole. It feels like a 90% match."
The SVM (The Calculator): Takes that "feeling" and runs it through a strict, mathematical rulebook. It says, "Okay, based on the geometry and the distance of these features, I am 100% sure this is a Macular Hole."

By combining the intuition of the AI artist with the precision of the math calculator, they created a super-team.

The Results: Who Won?

The researchers ran a race with these three models:

VGG-16 (The Old Detective): Finished with 83% accuracy.
ViT (The Modern Artist): Finished with 88% accuracy.
ViT-SVM (The Super-Team): Finished with 94% accuracy!

The hybrid model was the clear winner. It was especially good at spotting "Normal" eyes and "Diabetic Retinopathy" (getting those perfect 100% scores in the test), while the other models sometimes got confused.

Why Does This Matter?

Think of this like a medical triage system.

If a doctor has to look at 1,000 scans a day, they might miss a small hole in the 500th scan because they are tired.
If this ViT-SVM robot does the first pass, it can flag the 500th scan as "High Risk" with 94% confidence. The doctor then only needs to double-check the ones the robot is unsure about.

The Bottom Line

This paper shows that by mixing the best parts of two different types of AI (one that sees patterns like a human, and one that calculates like a machine), we can create a tool that helps doctors catch eye diseases earlier. Early detection means patients can get treatment before they lose their sight.

In short: Old AI is okay, new AI is good, but mixing them together makes the best doctor's assistant.

1. Problem Statement

The paper addresses the critical need for early and accurate detection of three prevalent retinal diseases that lead to partial or complete vision loss:

Diabetic Retinopathy (DR): A microvascular complication of diabetes.
Central Serous Retinopathy (CSR): Caused by serous detachment in the neurosensory retina.
Macular Holes (MH): Openings in the foveal center of the retina.

While Optical Coherence Tomography (OCT) scans are the standard for diagnosis, manual analysis by ophthalmologists is time-consuming and prone to human error. Existing automated solutions often rely on Convolutional Neural Networks (CNNs) or standard Machine Learning (ML) models, which may struggle with the high dimensionality of OCT data, computational costs, or performance trade-offs when model sizes are reduced. The authors aim to develop a hybrid architecture that leverages the strengths of both Deep Learning (feature extraction) and traditional Machine Learning (classification efficiency).

2. Methodology

The study proposes a novel ViT-SVM Hybrid Architecture and compares it against two baseline models: a standard Vision Transformer (ViT) and a pre-trained VGG-16.

A. Dataset and Pre-processing

Source: A publicly available dataset from the University of Waterloo (Lakshminarayanan et al.) containing 517 OCT-B scans.
Classes:
1. Normal Retina (206 images)
2. Diabetic Retinopathy (107 images)
3. Macular Holes (102 images)
4. Central Serous Retinopathy (102 images)
Pre-processing: Images were resized to 256x256. Data augmentation was applied using Keras ImageDataGenerator, including random vertical and horizontal flipping, with a batch size of 8.

B. Model Architectures

Baseline 1: VGG-16
- A classical 16-layer CNN pre-trained on ImageNet.
- Modified by removing the original classification head and adding a Flatten layer, a Dropout layer (0.5), and a Dense layer with Softmax activation.
Baseline 2: Vision Transformer (ViT)
- Utilized the ViT-B32 (Base model with 32x32 patch resolution).
- Images are split into patches, embedded, and processed through Transformer Encoder blocks (Multi-Head Self-Attention, Layer Norm, MLP).
- The output is passed through a custom head: Flatten $\rightarrow$ Dropout (0.5) $\rightarrow$ Dense (4 neurons, Softmax).
Proposed Model: ViT-SVM Hybrid
- Feature Extraction: Uses the ViT-B32 backbone to extract high-level features from the OCT images.
- Feature Transformation: The output of the ViT is passed through two Dense layers (64 neurons and 4 neurons).
- Classification: Instead of a standard Softmax layer, the features are fed into a Support Vector Machine (SVM) classifier.
- Configuration:
  - Loss Function: Squared Hinge (standard for SVM).
  - Regularization: L2 Regularization (factor 0.01).
  - Optimizer: Adam.
  - Activation: Softmax on the final dense layer for multi-class probability.

C. Implementation Details

Framework: TensorFlow 2 (Keras API).
Hardware: AMD Ryzen 5 3600 CPU, 16GB RAM, Nvidia GTX 1660 Ti GPU.
Training: 50 epochs using Categorical Cross-Entropy (for baselines) and Squared Hinge (for the hybrid).

3. Key Contributions

Novel Hybrid Architecture: The paper introduces a specific integration of Vision Transformers (ViT) for feature extraction and Support Vector Machines (SVM) for classification. This combines the global attention capabilities of Transformers with the memory efficiency and robust decision boundaries of SVMs in high-dimensional spaces.
Performance Benchmarking: It provides a rigorous comparison between a pure Deep Learning approach (ViT), a classical CNN (VGG-16), and the proposed hybrid model on a specific medical imaging task (OCT scans).
Efficiency vs. Accuracy: The study demonstrates that hybridizing a transformer with a traditional ML classifier can yield higher accuracy than using the transformer alone, potentially offering a more computationally efficient solution for medical diagnostics.

4. Results

The models were evaluated on Precision, Recall, and Accuracy across the four classes.

Model	Overall Accuracy	Key Observations
VGG-16	83%	Performed the poorest. Struggled with Macular Holes (Precision 0.45) and Central Serous Retinopathy (Precision 0.64).
ViT (Base)	88%	Improved significantly over VGG-16. Achieved perfect recall for Normal retina and high precision for DR and CSR.
ViT-SVM (Hybrid)	94%	Best Performance. Outperformed ViT by 6% and VGG-16 by 11%. Achieved 100% Precision and Recall for Normal and Diabetic Retinopathy classes.

Confusion Matrix Analysis: The ViT-SVM model showed the fewest misclassifications, particularly excelling in distinguishing between the four distinct retinal conditions.
Metric Breakdown:
- ViT-SVM: Perfect scores (1.00) for Normal and DR. 0.89 Precision/0.80 Recall for CSR. 0.82 Precision/0.90 Recall for Macular Holes.

5. Significance and Conclusion

Superiority of Hybrid Approach: The results confirm that combining the attention mechanisms of Vision Transformers with the decision-making power of SVMs creates a more robust classifier for medical imaging than using either component in isolation or relying solely on CNNs.
Clinical Relevance: Achieving 94% accuracy on a relatively small dataset (517 images) suggests the model is highly effective for early detection, which is crucial for preventing vision loss in patients with DR, CSR, and Macular Holes.
Future Directions: The authors suggest that while the current model is effective, future work should focus on fine-tuning the architecture further and expanding the dataset size to enhance generalizability and performance across diverse patient demographics.

In summary, this paper successfully validates that a ViT-SVM hybrid architecture is a state-of-the-art approach for classifying retinal OCT scans, offering a significant accuracy boost over traditional CNNs and standalone Transformer models.