Artificial Intelligent Systems and Machine Learning

Asha Gowda Karegowda ¹, Sudeshna Bhattacharyya ¹, M. A. Jayaram ¹, A. S. Manjunath ²

Affiliations
1 Dept. of MCA, Siddaganga Institute of Technology, Tumkur, Karnataka, IN
2 Dept. of Computer Science, Siddaganga Institute of Technology, Tumkur, Karnataka, IN

Abstract

Exudates are one of the primary signs of diabetic retinopathy, which is the main cause of blindness and can be prevented with an early screening process. In this paper, KNNFP and WKNNFP classifiers have been used for automatic exudates detection. The publicly available diabetic retinopathy dataset DIARETDB1 has been used in the evaluation process. The RGB image is converted to HIS color space. The median filter is applied to intensity image of HIS for removal of noise followed by Contrast-Limited Adaptive Histogram Equalization to achieve uniform illumination. Further the optic disk is eliminated since optic disk has properties similar to exudates,which may cause the hindrance with exudates detection. Five pixel level features were selected as input for classfication of exudates and non-exudates pixels: hue from hue image and intensity, mean intensity, standard deviation of intensity and distance between mean of optic disk pixels and pixels of exudates and non-exudates extracted from the preprocessed Intensity image. KNNFP and WKNNFP classifiers have been experimented using two distance measures namely Euclidean distance and Manhattan distance. Investigation reveals that the performance of KNNFP using Euclidean distance is superior when compared to KNNFP using Manhattan distance. WKKNFP has been experimented using three attribute weight assignment methods:Relief, information gain and Gain ratio.Compared to KNNFP, there is substantial improvement of WKKFP performance by assigning the feature weight Gain ratio and Relief method. The classfication accuracy of WKNNFP is found to be 97.50% compared to classfication accuracy of 96.67% with KNNFP classifier.

Keywords

Diabetic Retinopathy, KNNFP and WKNNFP, Image Preprocessing, Exudates.

Full Text

     

D. Poornima ¹, Asha Gowda Karegowda ²

Affiliations
1 Department of BCA, RRIT, Bangalore, IN
2 Department of MCA, SIT, Tumkur, IN

Abstract

A Diffuse goiter is usually an auto immune disease where the thyroid gland situated in the neck becomes enlarged. It is one of the most common thyroid disorders. In this paper, a diagnostic classifier system is designed to diagnose diffuse goiter using Thyroid Ultrasound (TUS) images.  Normal and goiter TUS images are pre-processed using Matlab software to extract Region of Interest (ROI). Relevant features were extracted from ROI. These features are validated by classifier system using different machine learning techniques- Naïve Bayes, Multilayer Perceptron (MLP), Radial Basis Function (RBF), Instance Based K-nearest neighbor (IBK) and J-48. The performance evaluation of the developed system is estimated using classification accuracy, time taken to build the model, Mean Absolute Error (MAE), Kappa Statistic and ROC Area. The obtained classification accuracies of different classifiers are in the range 88%-93%, 85%-91% and 83%-90% using 3-fold, 10-fold and hold-out (60%-40%) methods respectively. The results are very promising, indicating selected features were robust for diagnosis of diffuse goiter and also demonstrating the utility of machine learning techniques in medical domain.

Keywords

Machine Learning, Image Processing, Diffuse Goiter, Pre-Processing, Feature Extraction, Classification.

Full Text