Figure 4

Mandibular tomography completion results. The 2D results of the mandibular tomography completion were extracted and are presented as the perfect result (a) and the failed result (b). Each set of images contains the precompletion image, the image removal of the mask area, and the completed image. In (a1), we remove the lesion area, and we obtain the completion in (a2). (a2) shows that the completion can combine the generated area with the natural area. Therefore, although many of the completion results were good or perfect, some obviously failed, such as (b1,b2). (c) Shows the completion of the 3D image sequence for images from a real case. Due to the excessive number of images, only one of every 28 images in the lesion area is displayed. First, we use 2D completion initialized with a random input. Although only a few images fail, it is difficult to obtain all perfect results due to the many layers of 3D images. (c1) Shows a type of failure that is to complete the pattern with a malposition of the natural bone. (c2) Shows another type of failure in which part of the jawbone is deficient due to completion. To discard the poor results generated in the completion process, we establish the MOON model. In the first step of the MOON, we generate multiple completion results and remove the failed results using the image loss value. However, we still find some mistakes, such as (c3). Although the image looks similar to mandibular tomography, it should not have a tooth image in the lower part of the mandible, i.e., the layer is wrong. In the second step of the MOON, we calculate the semantic loss of each candidate image of the current tomography and its adjacent layer and obtain the final completed image.