PDR, CLAHE, SEGMENTASI, FITUR FRAKTAL, CLASSIFIER

• Main Author: Hanidar Ma'ruf, Afif
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/55923
• Institution: Institut Teknologi Bandung

Proliferative diabetic retinopathy, or PDR for short, is a disease caused by diabetes mellitus. This disease attacks the eye, especially the retina. If not treated immediately, patients can experience decreased ability to see or even total blindness. Doctors use retinal images to diagnose PDR. However, there are often differences in diagnoses between doctors, so CAD (Computer Aided Detection) was developed to assist doctors in diagnosing PDR. Many algorithms related to PDR detection have been developed. However, the majority of these algorithms use retinal images with good and stable quality as their input images. Even though not all hospitals have equipment that supports photographing retinal images. The retina dataset used, namely the Noor ul Huda dataset, has variations in image quality in lighting, sharpness and contrast. This final project proposes a new method of PDR detection that is able to adapt to variations in retinal image quality. This method consists of preprocessing in the form of CLAHE and median filter. As a result, preprocessing is able to significantly change the image quality and uniform the quality distribution. Segmentation is done by 2 types of methods, namely line detector based on simple line detector and Mlvessel algorithm based on wavelet filter. Extraction of fractal features and red lesions were the main features in the classification. Classification was carried out using 3 types of classifiers, namely L1 and L2 logistic regression and SVM. The combination of features analysis performed by ranking individual AUCs resulted in the highest performance in the line detector method without enhancement with features: 4 coefficients of fractal dimensions (box dimensions with and without skeletonization, information dimensions and coercion without skeletonization), 6 serial data of fractal measurements on all types and features red lesion. The highest performance was achieved using L2 logistic regression with AUC value of 0.9392 ± 0.003, sensitivity 94.51% and specificity 81.18%. Classification errors occur in non-PDR to PDR and PDR to non-PDR classes. Classification errors can occur due to image quality that is below the overall average quality and unbalanced data for each class.