INPAINTING OF DENTAL �PANORAMIC TOMOGRAPHY �VIA DEEP LEARNING METHOD

The tradition of image inpainting has existed for a long time; it is used to correct old and corrupted images. In recent times, progress in deep learning allows artificial neural networks to perform inpainting on clinical images to reduce image artifacts. In this paper, we demonstrated how various neural network models could perform inpainting on a dental panoramic tomography that was taken by using cone-beam computed tomography (CBCT). Experiments were done to compare the output of three different artificial neural network models: shallow convolutional autoencoder, deep convolutional autoencoder, and U-Net architecture. The dataset was taken from an open online dataset provided by Noor Medical Imaging Center. Qualitative assessment of the output shows that the U-net model reproduces the best output images with minimal blurriness. This result is also supported by the quantitative measurement, which shows that the U-net model has the smallest mean squared root error and the highest structural similarity index measure. The experiment results give an early indication that it is feasible to use U-Net to fix and reduce any image artifact that occurs in dental panoramic tomography