Automatic Screening of Diabetic Retinopathy at Early Stage using Colored Fundus Images

Abstract

The research work in this thesis is about an automated diagnosis of diabetic retinopathy (DR) for screening of DR patients at an early stage by classifying both red lesions using color fundus images, through a deep convolutional neural network (DCNN). DR is an asymptotic and vision-threatening retinal complication among working-age adults. Computer-aided diagnosis (CAD) is useful in screening DR patients to assist ophthalmologists to prevent blindness. Researchers have focused on this area by proposing many automated systems to diagnose DR. A comprehensive literature search is conducted during this research. Early-stage diagnosis is challenging due to poor representation of less discriminative and small-sized red lesions of DR. Traditional handcrafted based methods are not usually recommended for clinical trials due to their various limitations. Modern DCNNs are popular to solve various computer vision problems accurately. However, the requirement of huge training data restricts training deep models from scratch to solve the problems in medical imaging. Transfer learning and fine-tuning are helpful alternatives, but overfitting and poor performance further demand the architectural amendments for effective use of the deep CNN models in medical imaging with smaller datasets. Various pre-trained CNN models (AlexNet, VGG16, GoogLeNet, Inception-v3, ResNet50, and DenseNet) are fine-tuned, on the augmented set of 200 × 200 small image patches from the lesion-level annotated images of e-Ophtha_MA dataset, through transfer learning and hyper-parameter tuning to analyze their performance to classify both red lesions of DR. The architecture of a high-performing deep ResNet50 model is selected to fine-tune further. Its baseline architecture is modified by introducing; i) the reinforced skip connections, ii) a Global Max pooling layer, and iii) the Sum-of-Squared Error (SSE) Loss function. The suggested modifications are robust and straightforward. Such alterations may help to enhance various CNN architectures to apply in other imaging domains with small datasets. The proposed framework is evaluated on five publically available datasets: e Ophtha_MA, DiaRetDB1 v2.1, ROC, IDRiD, and Messidor by computing many xi performance metrics. The highest scores (0.9851, 0.991, 0.991, 0.991, 0.991, 0.9939, 0.0029, 0.9879, and 0.9879) of the sensitivity, specificity, AUC, accuracy, precision, F1-score, false-positive rate, Matthews’s correlation coefficient, and kappa coefficient are obtained on unseen test instances from e-Ophtha_MA for DR detection, respectively. The cross-validation results obtained by the proposed method on Messidor and IDRiD datasets at the image level are also better than state-of-the-art techniques. The achieved results are promising and demonstrate the effectiveness of a suggested architecture. For qualitative assessments, the gradient of class activations mapping (Grad-CAM) is computed to visualize the decision for each classified instance. This visual interpretation gives more satisfaction to using the proposed framework for clinical validations. Due to performance, simplicity, and robustness, the suggested model is suitable for diagnosing DR at an early stage for the screening of DR patients. It would help to develop a reliable health care system.