DESIGN AND EVALUATION OF POST-PROCESSING METHODS BASED ON THE RANDOM FOREST CLASSIFIER TO DECREASE FALSE POSITIVES IN MICROANEURYSM SEGMENTATION USING A U-NET MULTIRESOLUTION MODEL

• Main Author: Siti Sarah, Ines
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/80972
• Institution: Institut Teknologi Bandung

Microaneurysms (MAs), which are swellings in the blood vessels visible on the fundus of the eye, are one of the indicators for the early detection of diabetic retinopathy (DR), a complication of diabetes that causes a decrease in patients' visual ability. In previous research on MA detection in the IDRiD dataset (Ramadiastri, 2022), a MA detection model was developed using a multiresolution U-Net model, with a post-processing method of majority hard voting by 3 resolutions (640x640 px, 960x960 px, and 1280x1280 px). It was found that the precision value of the segmentation method was still low. The main objective of this study is to design a new post-processing method to improve the precision and F- score values from the Ramadiastri (2022) study. Three post-processing methods based on MA feature classification with a Random Forest Classifier were developed in this study. Methods A and B each perform MA feature classification one level before and after hard voting of the 3 resolutions, while method C performs MA feature classification with a 2-level classifier after hard voting of the 3 resolutions. Methods A, B, and C respectively provided reductions in False Positive (FP) by 70.83%, 68.193%, and 43.35%, but reduced True Positive (TP) by 21.580%, 18.989%, and 5.972%, while simultaneously adding False Negative (FN) by 39.014%, 34.497%, and 13.450%. Method C was considered to provide the best trade-off with precision, recall, and F-score of 0.458±0.033, 0.574±0.032, and 0.509±0.028 respectively. The values of shape features (major axis length, minor axis length, area) tended to increase significantly when reviewed between the True Negative (TN) and FN groups as well as between the FN and TP groups. Two-level classification successfully converted FN to TN, as indicated by the largest absolute decrease in Rank Biserial Correlation (RBC) and Common Language Effect Size (CLES) from 0.771 and 0.885 (Method B) to 0.438 and 0.719 (Method C).