Retinal Disease Classification from Fundus Images using CNN Transfer Learning

1. Problem Statement

Retinal diseases (e.g., diabetic retinopathy, glaucoma, AMD) are leading causes of preventable blindness, particularly in regions with limited access to ophthalmological care. Traditional diagnosis relies on manual fundus image examination by specialists, which is time-consuming, costly, and prone to inter-observer variability. While automated Computer-Aided Diagnosis (CAD) systems exist, challenges remain regarding:

• Data Scarcity: Limited annotated medical datasets compared to natural image datasets.
• Class Imbalance: Medical datasets often contain significantly more healthy samples than diseased ones, leading to biased models.
• Generalization: Models often struggle to detect subtle pathological patterns or generalize across different imaging conditions.

The study aims to develop a reproducible deep learning framework for binary risk classification (Normal vs. Diseased) to assist in early screening and clinical triage.

2. Methodology

The research employs a comparative experimental design using a publicly available Kaggle dataset containing 3,200 retinal fundus images. The dataset is split into Training (1,920), Validation (640), and Test (640) sets. The task is defined as a binary classification problem based on a "Disease Risk" label.

Two distinct experimental approaches were implemented:

A. Experiment 1: Baseline CNN (From Scratch)

• Architecture: A lightweight, custom Convolutional Neural Network (CNN) with two Conv2D layers (32 and 64 filters), MaxPooling, a Flatten layer, and a Dense output layer (Sigmoid activation).
• Preprocessing: Images resized to 64×64 pixels; pixel values rescaled to [0, 1].
• Training: No data augmentation was applied. Trained for 10 epochs using the Adam optimizer and Binary Cross-Entropy loss. No class weighting was used.
• Goal: To establish a baseline performance level using a simple architecture and low-resolution inputs.

B. Experiment 2: Transfer Learning (VGG16)

• Architecture: Utilized a pretrained VGG16 backbone (ImageNet weights, top removed) frozen during training, followed by a custom classification head (Flatten, Dense 500, Dense 100, Output with Softmax).
• Preprocessing: Images resized to 254×254 pixels; pixel values normalized to [0, 1].
• Data Augmentation: Applied horizontal and vertical flipping to increase dataset diversity and reduce overfitting.
• Imbalance Handling: Class weights were computed and applied during training to mitigate bias toward the majority (Normal) class.
• Training: Trained for 5 epochs using Adam optimizer, Categorical Cross-Entropy loss, and a learning rate of 0.01.

C. Evaluation Metrics

Performance was assessed using Accuracy, Precision, Recall, F1-score, Confusion Matrices, and Receiver Operating Characteristic (ROC) curves with Area Under the Curve (AUC) analysis.

3. Key Results

The study demonstrates a significant performance gap between training a model from scratch versus using transfer learning.

Metric	Baseline CNN (64x64)	VGG16 Transfer Learning (254x254)
Test Accuracy	83.1%	90.8%
Weighted F1-Score	Not explicitly stated	0.90
AUC (Class 0: Normal)	~0.81	0.81
AUC (Class 1: Diseased)	~0.81	0.81
Recall (Diseased)	Not explicitly stated	0.65
Recall (Normal)	Not explicitly stated	0.98
Precision (Diseased)	Not explicitly stated	0.88

• Performance Gain: The VGG16 model improved accuracy by 7.7 percentage points over the baseline.
• Class Specifics: The transfer learning model achieved high specificity for normal cases (98% recall) but struggled with the minority class, correctly identifying only 65% of diseased cases.
• ROC Analysis: Both models showed AUCs around 0.81, indicating that while the baseline had decent discriminative power, the transfer learning model provided superior overall feature representation and stability.

4. Key Contributions

1. Reproducibility: The study emphasizes strict methodological transparency, reporting results that correspond exactly to executed code runs without inferred metrics.
2. Comparative Analysis: It provides a clear empirical comparison between a shallow, low-resolution CNN and a deep, high-resolution transfer learning approach, quantifying the benefits of pre-trained features in medical imaging.
3. Imbalance Mitigation: It demonstrates the application of class weighting in a transfer learning context to address dataset skew, though it acknowledges this alone is insufficient for perfect minority class detection.
4. Binary Screening Focus: The study validates the utility of binary "Risk" classification as a viable first-line screening tool for clinical referral, rather than attempting complex multi-class diagnosis immediately.

5. Significance and Discussion

• Effectiveness of Transfer Learning: The results confirm that pre-trained architectures (VGG16) significantly outperform models trained from scratch on small medical datasets. The ability of ImageNet-pretrained filters to detect low-level edges and textures translates well to retinal vascular patterns, even with domain differences.
• Resolution Impact: Increasing input resolution from 64×64 to 254×254 allowed the model to capture subtle pathological lesions that were lost in the lower-resolution baseline.
• Clinical Implications:
  • Strengths: The high recall for normal cases (0.98) suggests the model is excellent at ruling out disease, reducing unnecessary referrals (high specificity).
  • Limitations: The recall for diseased cases (0.65) indicates that 35% of pathological cases were missed. In a screening context, missing diseased cases is critical. The authors note that while overall accuracy is high, the model requires optimization (e.g., threshold tuning, advanced loss functions like Focal Loss) to improve sensitivity for the minority class.
• Future Directions: The paper outlines a roadmap for improvement, including fine-tuning deeper layers of the VGG16 backbone, employing advanced data augmentation (GANs), optimizing decision thresholds, and integrating Explainable AI (Grad-CAM) to build clinician trust.

Conclusion:
The paper successfully demonstrates that transfer learning with VGG16 is a robust approach for automated retinal disease screening, achieving 90.8% accuracy. However, the persistent challenge of low recall for the diseased class highlights that while AI can assist in triage, further optimization is required before such systems can be reliably deployed as standalone diagnostic tools in clinical practice.