Multi-label Image Classification for Ocular Disease Diagnosis Using K-fold Cross-Validation on the ODIR-5K Dataset

Ocular diseases represent a significant global health concern. With the prevalence of such diseases rising, innovative diagnostic solutions are essential. This study presents a multilabel image classification model for diagnosing various eye fundus diseases using the ODIR-5K dataset. By transforming disease classes into binary vectors, this approach surpasses multi-class methods in modularity and versatility. The application of Kfold cross-validation in our models, particularly InceptionV3 and InceptionV3-ResNet50V2, yielded notable results, with precision rates of 0.96 and recall rates of 0.88, and an F1-score of 0.92, respectively, surpassing competing classifiers. This development offers more nuanced and accurate disease identification and lays the groundwork for integrating this model into an accessible, open-source web framework to benefit more people who suffer from ocular diseases. Such an advancement holds significant promise in improving ocular disease diagnosis, exemplifying the potential of machine learning in enhancing healthcare diagnostics. The code is available at https://github.com/andreivannuabc/ odir-5k-37-classes.