Deep Learning Framework Design for Diabetic Retinopathy Abnormalities Classification

Authors in this paper proposes an anomaly classification method based on a convolution neural network using the InceptionResnet-V2 deep neural network framework. The DR anomaly classification achieves excellent accuracy of 97.04%, as well as a sensitivity value of 95.10%, a specificity of 98.99%, and precision value of 99% for the MESSIDOR reference dataset, further with IDRiD dataset accuracy of 98.01%, and a sensitivity value of 97.06%, a specificity of 98.99%, and an precision value of 99%. The proposed method provides a minimum cross-entropy loss of 0.351, consuming a time of 15 minutes and 31 seconds. Significant performance improvement is observed for the other IDRID dataset used for the feasibility study of the proposed method compared to other mainstream models. The proposed method outperforms other state-of-the-art classification algorithms and provides a maximum accuracy improvement of 9.89% compared to the recent benchmark approach.