Jurnal Elkomika (Oct 2022)
Perbandingan Deteksi Letak Polip pada Citra Colonoscopy menggunakan CNN dengan Arsitektur RetinaNet
Abstract
ABSTRAK Penyakit kanker kolorektal diawali munculnya polip pada usus besar yang dapat berubah menjadi tumor ganas dan menimbulkan kanker. Sehingga diperlukan screening terhadap usus besar menggunakan colonoscopy. Menurut penelitian sekitar 26% polip terlewat saat prosedur colonoscopy. Pada penelitian ini dilakukan implementasi Convolutional Neural Network (CNN) dengan arsitektur RetinaNet untuk mendeteksi letak polip pada citra colonoscopy. Perbandingan dilakukan pada 3 jenis arsitektur yaitu ResNet-50, ResNet-101, dan ResNet-152 sebagai backbone pada arsitektur RetinaNet. Model yang terbaik berdasarkan metrik Intersection over Union (IoU) adalah model RetinaNet (Backbone = ResNet-50) tanpa data augmentation dengan nilai 0.8415. Sedangkan model yang terbaik berdasarkan metrik Average Precision (AP) adalah RetinaNet (Backbone = ResNet-101) dengan data augmentation dengan nilai AP25 = 0.9308, AP50 =0.9039, AP75 = 0.6985. Kata kunci: polip, colonoscopy, Convolutional Neural Network (CNN), RetinaNet  ABSTRACT Colorectal cancer always begins with the appearance of polyps in the colon which can turn into malignant tumors and cause cancer. Therefore, it is necessary to screen the large intestine using colonoscopy. However, according to studies, about 26% of polyps are missed during colonoscopy procedures. In this study, a Convolutional Neural Network (CNN) with RetinaNet architecture was implemented to detect the location of polyps in colonoscopy images. Comparisons were made on 3 types of architecture, namely ResNet-50, ResNet-101, and ResNet-152. From the evaluation results, the best model based on the Intersection over Union (IoU) metric is the RetinaNet model (Backbone = ResNet-50) without augmentation data with a value of 0.8415. While the best model based on the Average Precision (AP) metric is RetinaNet (Backbone = ResNet-101) with data augmentation with values AP25 = 0.9308, AP50 = 0.9039, AP75 = 0.6985. Keywords: polyp, colonoscopy, Convolutional Neural Network (CNN), RetinaNet
Keywords