Hybrid Neuro-Fuzzy Approaches for Abnormality Detection in Retinal Images

Abnormality detection in human retinal images is a challenging task. Soft computing techniques such as neural approaches and fuzzy approaches are widely used for these applications. However, there are significant drawbacks associated with these approaches. Artificial Neural Networks (ANN) yield high accuracy only when the training data is sufficiently large and accurate. On the other hand, fuzzy approaches are quite accurate but require significant computational time. Hence, a combination of neural and fuzzy approach is tested in this work which yields high accuracy within a reasonable time. The neuro-fuzzy model used in this work is Adaptive Neuro-Fuzzy Inference Systems (ANFIS) which possess the benefits of neural approaches and fuzzy approaches. The applicability of these techniques is explored in the context of categorizing the normal and abnormal retinal images. The performance of the classifiers is analyzed in terms of sensitivity, specificity, and classification accuracy and convergence time. Representatives from neural approaches and fuzzy approaches are also implemented for comparative analysis. The neural and fuzzy approach used in this work is Kohonen Neural Network and Fuzzy C-Means (FCM), respectively. Experimental analysis suggests promising results for the hybrid approach.