Authors - A Aruna kumari, Sri Vishnu Prabhu Gudavalli, Tamminana Visweswari Abstract - This project presents a hybrid architecture, which combines sequential and non-sequential models, to detect diabetic retinopathy (DR) in retinal fundus image data using deep learning. The model has a non-sequential backbone feature extraction layer, a pre-trained ResNet50, built using the Functional API of Keras to allow flexibility and skip connections. On top of that, a functional custom sequential classifier head is added, comprising of, e.g., GlobalAveragePooling2D, Dense, and Dropout layers, also constructed in a functional manner to ensure the architecture is coherent. The classifier is used to make a binary prediction: DR or not. The model was trained using the EyePACS data using different techniques of image preprocessing, including resizing, normalization, and augmentation. The resulting architecture is completely accurate and exhibits good generalization on unseen data, which indicates the effectiveness of transfer learning in combination with deep sequential classifiers. This bifurcated architecture is understandable and performance-wise attractive to be used in clinical decision support in diabetic retinopathy screening.
Open Zoom