Cataract Disease Detection Using VGG 19 Deep Learning Architecture

The study develops a DL method that automatically detects cataracts from fundus images using a VGG19-based CNN. The Ocular Disease Recognition ODIR5K dataset was used for data collecting a balanced collection of 589 cataract images and 500 fundus images devoid of any disease was created. The image was loaded, resized and normalized, then the data set was augmented to raise its strength. Binary classification drew on a VGG19 model of transfer learning. Frozen convolutional layers were included along with a unique classification head. Test-to-training set ratios of80:20 helped the model to be trained. Over-fitting was avoided in part via early stopping, model checkpoint decisions, Adam optimizer, and binary cross-entropy loss function. After 15 epochs of training, the model reached an accuracy of 98.17% in both cases. The model generalized well to new data by obtaining an F1 score on its labels when tested. With a 98.17% accuracy, the model precisely identified cataracts at 96%. Still, its recall was below average just 100% successful. The high sensitivity and specificity of the model were shown by the confusion matrix. Moreover, random test sample eye inspection revealed the broad general resilience of the approach. This work uses VGG19 transfer learning in a cataract diagnosis framework. Moreover, constant application of the approach not only guarantees great precision but also produces stability. Future studies will probably help ophthalmologists in identifying cataracts and improving diagnosis efficiency for patients as this technique is included in clinical procedures. Combining this approach with other imaging modalities should specifically help to improve clinical results as well. Plans call for studying more complex deep learning architectures and including other eye diseases. This work focuses attention on world health policies and tackles health in low-resource environments.