A Predictive Deep Learning Models For Automated Retinal Aberration Detection

Abstract

Retinal fundus images play a significant role in the diagnosis and assessment of various ocular diseases by ophthalmologists. In recent years, extensive research has focused on the application of deep learning techniques to retinal fundus images for enabling early disease detection and facilitating timely treatment. The capability of deep learning models to rapidly analyze medical images and generate diagnostic results can support faster clinical decision-making and treatment planning. This study proposes a non-invasive approach for the early identification and management of multiple eye diseases using a Convolutional Neural Network (CNN).

The Retinal Fundus Multi-disease Image Dataset (RFMiD) was utilized in this research, comprising fundus images representing multiple ocular conditions, including Media Haze (MH), Optic Disc Cupping (ODC), Diabetic Retinopathy (DR), and healthy or normal images (WNL). Several image preprocessing and data preparation techniques were employed to enhance the effectiveness of the proposed models. These techniques include data augmentation, image cropping, resizing, dataset partitioning, conversion of images into numerical arrays, and one-hot encoding. CNN architectures were subsequently employed to automatically extract relevant and discriminative features from the color fundus images. The extracted features were then used to classify the images into their corresponding disease categories.

Three CNN architectures were developed and evaluated experimentally. Their performance was measured using commonly adopted statistical evaluation metrics, including accuracy, precision, recall, and F1-score. The experimental findings demonstrate that the 12-layer CNN achieved a validation accuracy of 89.81% and a testing accuracy of 88.72% after applying data augmentation. In comparison, the 20-layer CNN with augmented data achieved a higher validation accuracy of 90.34% and a testing accuracy of 89.59%. Although the 20-layer CNN produced better accuracy, the model exhibited signs of over fitting. Overall, the obtained results indicate that the proposed deep learning approach can effectively learn and differentiate between multiple ocular disease categories and healthy retinal images.

The primary contribution of this study is the development of an efficient and reliable deep learning-based diagnostic framework capable of simultaneously identifying multiple eye diseases from color retinal fundus images. The proposed approach demonstrates the potential of CNN-based models as a non-invasive and automated tool for supporting early ocular disease detection and clinical diagnosis.