Introduction
Retinal vessel segmentation is a critical diagnostic step for identifying pathologies such as Diabetic Retinopathy. While Convolutional Neural Networks (CNNs) are the gold standard for this task, their high computational and memory requirements often preclude their use in portable, battery-operated diagnostic devices. Handheld tools, such as binocular indirect ophthalmoscopes used in field surgery, typically rely on low-power hardware (e.g., FPGAs or mobile CPUs) that cannot support standard, "heavy" deep learning models. This paper presents a strategy to bridge this gap by creating a "lean" network that maintains clinical accuracy while significantly slashing computational costs.
Proposed Methodology
The authors propose a multi-stage pipeline that combines image preprocessing with a hybrid model simplification strategy. The process begins with Image Enhancement, using local histogram equalization on grayscale versions of retinal images to improve the contrast of vessels against the background.
The core of the innovation lies in the Hybrid Simplification of the CNN:
- Deterministic Ternary Quantization: This is applied to the Fully Connected Layers (FCLs). By restricting weights to just three values (-1, 0, +1) instead of 32-bit floating-point numbers, expensive multiplications are replaced with simple additions or subtractions.
- Network Pruning: This is applied to the Convolutional Layers (CLs). The algorithm identifies and removes "unimportant" connections (weights with low variance), reducing the overall parameter count.
To mitigate accuracy loss, the authors utilize an iterative "train-simplify-retrain" loop. After pruning or quantizing, the network is retrained to allow the remaining weights to compensate for the lost information. The model operates on 9x9 pixel patches extracted from the images, ensuring local feature focus.
Experimental Results
The model was validated using the STARE (STructured Analysis of the REtina) dataset with 5-fold cross-validation. The results demonstrate that high efficiency can be achieved with minimal impact on diagnostic performance:
- Weight Reduction: Approximately 60% of the weights in the convolutional layers were successfully pruned.
- Accuracy: The original complex CNN achieved an accuracy of 0.9617, while the simplified version maintained an accuracy of 0.9581—a drop of less than 0.4%.
- Clinical Metrics: The simplified model achieved a Sensitivity of 0.7599, a Specificity of 0.9757, and a high Area Under the Curve (AUC) of 0.97.
Conclusion
This research demonstrates that strategic pruning and quantization can enable the deployment of sophisticated retinal diagnostic tools on low-power, portable hardware. By reducing memory access energy consumption by a theoretical factor of 16 to 32, this approach paves the way for real-time, high-accuracy diagnostics in field and surgical environments. Future work will focus on hardware-level benchmarking and testing the model's generalization on lower-quality images from handheld cameras.
Summary updated