Machine Learning for the Real World: Applications and Insights

Diabetic retinopathy (DR) is a leading preventable cause of blindness worldwide.

It often has advanced progression prior to exhibiting significant symptoms.  Early detection and accurate grading of DR are important for treatment and also for the prevention of visual loss. The severity levels of No DR, Mild, Moderate, Severe, and Proliferative DR have been categorised for retinal fundus images for this analysis using a Convolutional Neural Network (CNN) model. This chapter investigates the application of CNN model to detect Diabetic retinopathy. Grad-CAM or gradient-weighted Class Activation Mapping will be applied thereafter to improve interpretability by highlighting clinically relevant features like haemorrhages, microaneurysms and exudates. Using preprocessing techniques such as scaling, normalisation, and augmentation improved the model's ability to generalise on both the EyePACS and Messidor datasets. The CNN produced a macro-averaged recall of 92%, accuracy of 93%, precision of 91%, and ROC-AUC of 0.95. In other words, dividing the remaining probabilities by No DR (23.9%), Mild (7.5%), Severe (6.5%) and Moderate (3.8%) respectively, while there was a single case of Proliferative DR that had a 58.3% confidence. To deploy it in real time, a simple web application with Flask was developed to get Grad-CAM overlays and predictions in seconds. The system is a viable option for early DR screening due to its accuracy, interpretability, and usability, particularly in low-resource settings and clinical settings.

Previous studies have initially demonstrated the effectiveness of deep learning-based systems for the detection and grading of diabetic retinopathy [1,6,11,15], which further encourages the adoption of CNN techniques in clinical screenings.