Anirban Mitra1, Priya Shankar Banerjee2, Sudipta Roy3, Somasis Roy4, Sanjit Kumar Setua4. 1. Department of Computer Science and Engineering, Calcutta University Technology Campus, JD-2, Sector-III, Salt Lake, Kolkata 700098, India; Department of Computer Science and Engineering, Academy of Technology, Adisaptagram 712121, West Bengal, India. 2. Department of Computer Science and Engineering, Academy of Technology, Adisaptagram 712121, West Bengal, India. 3. Department of Computer Science and Engineering, Calcutta University Technology Campus, JD-2, Sector-III, Salt Lake, Kolkata 700098, India; Mallinckrodt Institute of Radiology Department (MIR), Washington University School of Medicine, Campus Box 8225, 510 South Kingshighway Boulevard, Saint Louis, MO 63110-1076, USA. Electronic address: sudiptaroy01@yahoo.com. 4. Department of Computer Science and Engineering, Calcutta University Technology Campus, JD-2, Sector-III, Salt Lake, Kolkata 700098, India.
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.
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.