The region of interest localization for glaucoma analysis from retinal fundus image using deep learning

Abstract

Background and objectives: Retinal fundus image analysis without manual intervention has been rising as an imperative analytical approach for early detection of eye-related diseases such as glaucoma and diabetic retinopathy. For analysis and detection of Glaucoma and some other disease from retinal image, there is a significant role of predicting the bounding box coordinates of Optic Disc (OD) that acts as a Region of Interest (ROI).

Methods: We reframe ROI detection as a solitary regression predicament, from image pixel values to ROI coordinates including class probabilities. A Convolution Neural Network (CNN) has trained on full images to predict bounding boxes along with their analogous probabilities and confidence scores. The publically available MESSIDOR and Kaggle datasets have been used to train the network. We adopted various data augmentation techniques to amplify our dataset so that our network becomes less sensitive to noise. From a very high-level perspective, every image is divided into a 13 × 13 grid. Every grid cell envisages 5 bounding boxes along with the corresponding class probability and a confidence score. Before training, the network and the bounding box priors or anchors are initialized using k-means clustering on the original dataset using a distance metric based on Intersection of the Union (IOU) over ground-truth bounding boxes. During training in fact, a sum-squared loss function is used as the prediction's error function. Finally, Non-maximum suppression is applied by the proposed methodology to reach the concluding prediction.

Results: The following projected method accomplish an accuracy of 99.05% and 98.78% on the Kaggle and MESSIDOR test sets for ROI detection. Results of proposed methodology indicates that proposed network is able to perceive ROI in fundus images in 0.0045 s at 25 ms of latency, which is far better than the recent-time and using no handcrafted features.

Conclusions: The network predicts accurate results even on low-quality images without being biased towards any particular type of image. The network prepared to see more summed up depiction rather than past works in the field. Going by the results, our novel method has better diagnosis of eye diseases in the future in a faster and reliable way.

Keywords: Anchor Boxes; Batch Normalization; Convolution Neural Networks; Intersection over Union; K-means clustering; Leaky ReLU, Max Pooling; Non-maximum suppression; Optic Disc Localization.