Abstract
Diabetes is a chronic condition affecting millions of people worldwide. One of its major complications is diabetic retinopathy (DR), which is the most common cause of legal blindness in the developed world. Early screening and treatment of DR prevents vision deterioration; however, the recommendation of yearly screening is often not being met. Mobile screening centers can increase DR screening, but they are time and resource-intensive because a clinician is required to process the images. This process can be improved through computer-aided diagnosis, such as by integrating automated screening on smartphones. Here we explore the use of a SqueezeNet-based deep network trained on a fundus image dataset composed of over 88,000 retinal images for the purpose of computer-aided screening for diabetic retinopathy. The results of this neural network validated the viability of conducting automated mobile screening of diabetic retinopathy, such as on a smartphone platform.
Introduction
Diabetic Retinopathy (DR) is a severe complication of diabetes that damages retinal blood vessels. While early detection is critical for preventing vision loss, current screening processes are bottlenecked by the need for manual grading by trained clinicians. In many regions, the patient-to-specialist ratio makes universal yearly screening unfeasible.
While cloud-based AI solutions exist, they rely on stable internet connectivity and expensive server infrastructure. This research explores the feasibility of moving diagnostic tools directly onto smartphones. By utilizing compact neural network architectures, the authors demonstrate that high-performance DR screening can be achieved without the massive computational requirements of traditional deep learning ensembles.
Methodology and Data
The researchers employed the SqueezeNet architecture, prioritizing efficiency over raw parameter count. Unlike "heavy" models like ResNet or Inception, SqueezeNet uses "Fire Modules"—comprising 1x1 "squeeze" layers and a mix of 1x1 and 3x3 "expand" layers—to drastically reduce the number of parameters while maintaining representational power.
Preprocessing and Inference
Retinal images often suffer from inconsistent lighting and artifacts. The team implemented a robust preprocessing pipeline to standardize inputs:
- Otsu’s Method: Used to automatically crop black borders.
- Contrast Enhancement: Applied to highlight lesions such as microaneurysms and hemorrhages.
- 5-Crop Inference: During testing, predictions from the four corners and the center of the image were averaged to ensure stability.
Dataset
The model was trained and validated using the 2015 Kaggle Diabetic Retinopathy Detection dataset (EyePACS), consisting of 35,126 training and 53,576 test images. Images were graded on a 5-point clinical scale, ranging from no disease to proliferative DR.
Results
The compact model, containing only 807,013 parameters and occupying a mere 9.7 MB, achieved an overall accuracy of 83.6%. More significantly, it reached a Quadratic Weighted Kappa score of 0.76543. This performance would have placed the model in the top 20 of the original Kaggle competition (out of over 600 teams), despite many competitors using much larger ensembles.
Discussion and Future Work
The study proves that the "10MB threshold" for mobile deployment is achievable without sacrificing clinical relevance. A model of this size is easily distributable via standard app stores and can run locally on mid-range mobile hardware.
However, challenges remain for real-world deployment. Future work must focus on clinical validation with "in-the-wild" images captured directly by smartphone-attached lenses and incorporating patient-level analysis (comparing both eyes) to refine individual diagnoses.
Summary updated