Abstract
Premature new-borns are at risk for Retinopathy of Prematurity (ROP), a condition that can result in blindness. Those with low birth weights are more likely to experience this. Early detection and treatments are crucial to prevent blindness or severe visual impairment. Unfortunately, the current methods for diagnosing ROP are sometimes subjective, labourintensive, and require specialized knowledge, which delays the process of both diagnosis and treatment. Deep Learning (DL) algorithms show remarkable efficiency in a variety of medical imaging tasks, including the detection and classification of illnesses. The suggested hybrid Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) model, which offers a viable solution, combines the best aspects of the CNN-LSTM architectures. The CNN component efficiently extracts spatial properties from retinal images, whereas the LSTM component processes sequential image data across time to capture temporal dependencies. The diagnostic precision is further enhanced by the model’s increased ability to precisely analyse the progression of ROP. Blood vessel segmentation in fundus image is used to detect the abnormality effectively. Modified MultiResUNet model is used for segmentation of blood vessels. The robustness and accuracy of the suggested model are assessed using performance metrics such the F1-score, sensitivity, specificity, precision, and accuracy. Comparative studies using existing screening methods demonstrate that the proposed deep learning methodology is more effective in terms of accuracy of 97%.
Keywords: Convolutional Neural Network, Deep Learning, Long Short-term Memory, Retinopathy of Prematurity, Segmentation.