Implementation of Neural Network Classification for Diabetes Mellitus Prediction System through Iridology Image

Abstract

One alternative and a non-invasive method named iridology, has been developed to find more effective way of detecting diabetes mellitus. Iridology is the method of mapping the human organs, and it has corresponded in iris' zone. It can be used to detect damaged tissues, particularly in the pancreas where it holds the primary role of producing insulin. This study focuses on developing a non-invasive diabetes mellitus prediction system through an iris image using an image processing algorithm and neural network model. The processing starts with image enhancement using FFT filter and grayscaling, iris localization using Circular Hough Transform (CHT), and normalization using rubber sheet normalization. Segmentation on pancreas in iris image then resulted as followed, one ROI of right-eye image and two ROIs of left-eye image. The image database is collected with maximum of three images taken from 15 healthy subjects and 11 diabetes subjects, resulted in 201 data images. Feature extraction method that has been used is the Gabor filter, using the texture feature of the segmented iris image. The evaluation method we use for the system is the confusion matrix to obtain its accuracy and other parameters. Classification model of Feed-Forward Neural Network (FNN) is implemented to classify between diabetes and healthy subjects with the best results of accuracy number 95.74% and 92.57% for training and testing data respectively. The result shows that this system can be proposed as a complementary tool for therapeutic methods for diabetes prediction.