Perbandingan Metode Resnet50V2 dan Densenet201 Untuk Klasifikasi Penyakit Mulut Dengan Citra RGB

Abstract

Oral and dental diseases are common health problems that can lead to serious complications if not detected early. Conventional diagnosis through visual examination still has limitations, such as dependency on clinical experience and the complexity of visual variations in oral conditions. Therefore, an objective and automated approach is required to support oral disease classification. This study aims to compare the performance of two Convolutional Neural Network (CNN) architectures, namely ResNet50V2 and DenseNet201, for classifying oral diseases using RGB images. The dataset used in this study is a combination of three public datasets: Oral Disease Dataset, Oral Infection Dataset, and Teeth Dataset, consisting of six classes, namely calculus, gingivitis, caries, tooth discoloration, hypodontia, and mouth ulcer. The data were preprocessed through image resizing, data splitting with an 80:10:10 ratio, and data augmentation using Gaussian noise, motion blur, and downsampling techniques. Transfer learning was applied with two learning rate values (0.001 and 0.0001) and two loss functions, categorical crossentropy and focal loss, resulting in eight experimental scenarios. Model performance was evaluated using accuracy, loss, precision, recall, and F1-score. The experimental results indicate that DenseNet201 generally outperforms ResNet50V2, particularly when combined with focal loss and a learning rate of 0.0001, achieving more balanced and robust performance across all classes, including minority classes. These findings demonstrate that the dense connectivity mechanism of DenseNet201 is more effective in extracting complex features from RGB oral images, making it a more suitable architecture for oral disease classification.