Dental caries or tooth decay is damage to the hard tissues of the teeth that can occur in the enamel, dentin, and cementum areas. Panoramic radiography is a screening tool for tactile or visual examination of the oral cavity which is useful for further diagnosis and treatment. The process of segmentation of panoramic radiographs is a difficult process because there is no homogeneity between panoramic images with one another. Noise levels, vertebral column images, and low contrast are the main challenges in image processing. This study evaluates CNN to detect caries automatically on panoramic radiographs on a small dataset. The dataset consisted of manually cropped maxillary and mandibular premolars and molars. An augmentation strategy consisting of horizontal flip, vertical flip, and affine transformation is used to produce a wider variety of images. This study compares the architecture of non-pretrained and pretrained models consisting of 3-layer CNN, 3-layer CNN with batch normalization, ResNet18, and ResNeXt50 32×4d. Evaluation was carried out on 400 training data and 76 testing data. Combination of augmentation strategies and pre-trained ResNet18 and ResNeXt50 32×4d achieves high accuracy compared to other models.