OPTIMIZING HARD EXUDATE SEGMENTATION IN RETINAL IMAGES USING GRADIENT BOOSTING CLASSIFIER AND ACTIVE CONTOUR MODEL

• Main Author: Nihayatal Izza Sarinop, Ufi
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/78143
• Institution: Institut Teknologi Bandung

Diabetic retinopathy, one of microvascular complications caused by diabetes mellitus, ranks among the leading causes of blindness among the working-age population. Timely detection of diabetic retinopathy holds pivotal importance in mitigating the risk of blindness stemming from diabetic retinopathy. Manual detection, however, is both time-intensive and subjective. In the early stages, hard exudates emerge as a prominent clinical sign in retinal images. These exudates appear as sharp-edged, bright yellow spots, resulting from lipid leakage after retinal vessel rupture. Consequently, the development of an automated detection system emerges as a necessity, facilitating medical practitioners in diagnosing diabetic retinopathy. Among established methodologies, the segmentation of hard exudates in retinal images stands out as a promising approach. This study extends the work of Vasmaulidzra, who successfully trained a hard exudate segmentation model on IDRiD dataset using U-Net architecture. While the evaluation metrics demonstrate favorable overall results, there remains an extensive number of false positives. This study aims to address this concern comprehensively by implementing a post-processing approach, which involves the integration of adaptive active contour and gradient boosting classifier. To achieve this goal, a gradient boosting classifier is employed to address various aspects: mitigating false positive lesions, identifying lesions for further refinement through Adaptive Active Contour Model (ACM), and determining optimal parameter values for the ACM. These parameters includes the trajectory and rate of evolution. Additionally, a rule-based approach is employed to determine ACM parameters, including the type of ACM, scale factor, smoothing factor, and stopping criterion. The empirical findings of this study yield F-score, precision, recall, and accuracy metrics of 0.675 ± 0.120, 0.758 ± 0.092, 0.639 ± 0.170, and 0.994 ± 0.009, respectively. These figures, showcasing an average metric enhancement of 7.8% in F-score, 12.5% in precision, 1.8% in recall, and less than 1% in accuracy over the initial metric values, which exemplify the efficacy of the proposed methodology in optimizing hard exudate segmentation.