DR. SHRUTHY VAISHALI RAMESH

DR. PRASANNA VENKATESH RAMESH, Dr. R. RAMESH, Dr. SATHYAN PARTHASARATHY

Abstract

Purpose: A novel convolutional neural network (CNN) approach in detecting diabetic retinopathy (DR) was proposed, from true colour confocal fundus images, to overcome blackbox dilemma in artificial intelligence (AI). In addition to identification and classification, this tool intended to identify signs from trivial micro-aneurysms to catastrophic neo-vascularisation, due to customised annotations. Methods: 8000 high resolution confocal fundus images were utilised; 6600 images for training and 1400 images for testing. 114 signs pertaining to DR were annotated. You Only Look Once 5 algorithm was used for detection. The testing images were spilt into three groups: 400, 600 and 400 for three runs done, once every two months. Results: Tests showed consistent increments from 79.5% to 83% accuracy in predicting diagnosis, severity and intricate signs. Conclusion: With constant training via feedback mechanism, there was an upsurge in prediction accuracy, which helped overcome blackbox dilemma.

Full Text