Promax

A scanned cone-beam computed tomography (CBCT) of a completely edentulous mandible of a 68-year-old male patient was used as the morphological base for the FEM models (ProMax, Planmeca, Helsinki, Finland). The selection of the image data as the anatomical template was based on the patient's medical record status with age-appropriate health and bone status, representing no morphological and mineralization variabilities. The image data with pixel conditions of 651 × 651, at 96 kV, and with increment slices of 0.2 mm in thickness was then converted to DICOM format. Using established image processing software (Amira) the acquisition of cortical and cancellous mandibular bone architecture data was achieved by semi-automatic segmentation of coronary CT layers. The reticulation of point clouds (Delauney-triangulation) to three-dimensional polygon meshes produced morphologically identical sub-models of the cortical and cancellous mandible (Fig. 1).Fig.1

3-D edentulous mandible model with a two unsplinted interforaminal implants, b two splinted interforaminal implants, c four unsplinted interforaminal implants, d four splinted interforaminal implants.

All patients underwent standardized routine clinical evaluation after implant insertion and prosthetic restoration, with a follow-up every 3 months within the first year and every 12 months subsequently [16 (link)]. During follow-up examinations, peri-implant bone level, modified bleeding index (mBI), modified plaque index (mPI) [17 (link), 18 ], attached keratinized gingiva (AG), and pocket depth (PD) were measured. Mesial and distal peri-implant bone level changes were measured using the method described by Gomez-Roman et al. [19 , 20 (link)]. Implant success was evaluated according to the method described by Buser et al. [21 (link)]. Mean peri-implant bone level ± standard deviation was determined by examining the mesial and distal after 1, 3, and 5 years. A subdivision was done to distinguish between VP (VP +  = VP; VP- = no VP), irradiation (R +  = irradiation; R- = non-irradiation), and irradiation with or without VP (VP + R +  = irradiation and VP; VP- R +  = irradiation without VP). Routine standardized panoramic radiographs (trademark: Planmeca ProMax; type: ProMax 3D Max, Pro Face Med Series H23 120 kV) were regularly included in follow-up controls after 1, 2, 3, and 5 years. All patients involved in this study received a dental prophylaxis by trained personnel every 3 months.

The panoramic images were obtained using a Planmeca ProMax^® X-ray unit (Planmeca Oy, 00880 Helsinki, Finland), which was equipped with a digital sensor called Planmeca Dimax 3. The measurement was performed using the Planmeca Romexis^® software (Fadente distribution, Badalona, Barcelona, Spain, updated November 2023). Images were taken based on the manufacturer’s recommendations at 64–70 kV and 7–14 mA, depending on the patients’ gender and age. The recommended settings varied for adult females, small adult males, and large adult males. Well-trained doctoral students from the Oral Radiology Department conducted the measurement twice, with a two-month interval between each session. The intra-examiner reliability was calculated using the Kappa statistic [17 (link)]. Using the obtained OPGs, the level of the bone was measured from the cementoenamel junction of the second molar to the level of the alveolar crest for the study group before (Figure 1A) extraction and after extraction (Figure 1B).
Similarly, bone changes in the control group were measured at baseline (Figure 2A) and after a period of 3–6 months (Figure 2B).

In total, 888 dental panoramic radiographs were collected for this study from a private oral and maxillofacial radiology center in Tehran, Iran. A comprehensive sample from patients who visited the radiology center in June 2021 was used in this study, assuming that the number of images was sufficient to demonstrate the effects of SR in this exploratory study. Low-quality images (e.g., blurry or noisy images) were excluded (n = 31). All the images were taken using Planmeca ProMax (Planmeca, Helsinki, Finland). The device settings were 64–72 kVp, 6.3–12.5 mA, and 13.8–16 s exposure time. The images were exported to .jpg format with a size of 2943 × 1435. All the samples were anonymized before use in the study.
To assess whether deep learning and other conventional approaches were suitable for improving resolution, we downscaled all HR images by a factor of 4× to create LR images. Such generic downsampling has been previously employed to simulate LR images and allows high standardization and replicability [14 (link)]. The original HR images were considered as ground truth. Seventy percent of the images (n = 622) were selected as the training set. Of the remaining images, 50% were selected as a test set (n = 133), and the other 50% were selected as a validation set (n = 133).