Yolla Faradhilla, - and Agus Zainal Arifin, - and Nanik Suciati, - and Eha Renwi Astuti, - and Rarasmaya Indraswari, - and Monica Widiasri, - (2021) Residual Fully Convolutional Network for Mandibular Canal Segmentation. International Journal of Intelligent Engineering and Systems, 14 (6). pp. 208-219. ISSN 2185310X

	Text (ARTIKEL) 18. Residual Fully Convolutional Network for Mandibular Canal Segmentation.pdf Download (885kB)
	Text (TURNITIN) 18. 12% Residual Fully Convolutional Network for Mandibular Canal Segmentation.pdf Download (4MB)
	Text (VALIDASI) 18. Residual Fully Convolutional Network for Mandibular Canal Segmentation.pdf Download (414kB)
	Text (KARIL) 15.. Residual Fully Convolutional Network for Mandibular Canal Segmentation.pdf Download (135kB)

Abstract

In performing dental implant surgery on the mandible, it is necessary to measure the height and width between alveolar bone and mandibular canal. A system that is able to measure the height and width automatically is required since a manual measurement takes a long time. One of the important processes in this system is mandibular canal segmentation. The main problem in this segmentation process is the unbalanced numbers of data between object and background classes. This problem often led to misclassifications, especially at pixels on the object boundary. This research proposed a new architecture based on residual fully convolutional network (RFCN) by considering the loss values in the region and boundary of the mandibular canal segmentation. Dual auxiliary loss (DAL) functions are introduced to optimize RFCN, so that the network performs object segmentation better. For this research, DAL used focal loss to calculate the loss value in region and boundary to overcome the unbalanced class between the mandibular canal and the background. There are 2 datasets used for this research. The first dataset contains 200 images with mandibular canal and the second dataset contains 300 images consisting of 200 images with mandibular canal and 100 images without mandibular canal. We tested our network and compared it with state-of-the-art segmentation methods. The experiment showed that the proposed method outperforms all the comparing methods with Dice Similarity Score of 0.914 on the first dataset and 0.868 on the second dataset.

Actions (login required)

View Item