Mandibular Fractures

Primary oral osteoblasts (hOBs) were isolated from mandible bone fragments of n° 12 patients that underwent the surgical removal of lower third molars at the dental clinic of the G. D’Annunzio University. All patients signed an informed consent in accordance with the Declaration of Helsinki principles and according to the ethical standards of the Institutional Committee on Human Experimentation (reference number: BONEISTO N. 22 10 July 2021). Immediately after sampling, each bone fragment underwent three enzymatic digestions at 37 °C for 20, 30 and 60 min utilizing a solution consisting of collagenase type 1A (Sigma-Aldrich, St. Louis, MO, USA) and trypsin-EDTA 0.25% (Sigma-Aldrich) dissolved in Dulbecco’s Modified Eagle’s medium (DMEM, Corning, New York, NY, USA) at 10% fetal bovine serum (FBS, Gibco-Life Technologies, Monza, Italy). The solution obtained from the enzymatic digestion was centrifuged at 1200 rpm for 10 min. Then, the pellet obtained was transferred into a T25 culture flask with low-glucose (1 g/L) DMEM supplemented with 10% FBS, 1% antibiotics (100 µg/mL⁻¹ streptomycin and 100 IU/mL⁻¹ penicillin), and 1% L-glutamine to promote a final spontaneous migration of the cells. The isolated hOBs were cultured at 5% CO₂ and 37 °C to achieve their confluence to be used between the 3rd and the 5th passage upon the characterization by cytometric analysis. Following 10 days of culture, bone fragments were removed.

The sEMG examinations were conducted between 8 and 12 a.m. to minimize the influence of daily fluctuations of muscle activity. The electromyographic measurements were carried out in the same dental chair in a sitting position (the body perpendicular to the ground, the head resting on the chair’s headrest, and the lower limbs upright and arranged parallel). The height of the headrest was adjusted individually to set the head, neck, and torso of the subjects in a straight line.
Before placing the surface electrodes, the skin was cleaned with 90% ethanol solution to reduce skin impedance. Next, surface electrodes (Ag/AgCl with a diameter of 30 mm and a conductive surface of 16 mm (SORIMEX, Toruń, Poland) were placed bilaterally following the course of the muscle fibers of the temporalis anterior (TA), the superficial part of the masseter muscle (MM), the anterior bellies of the digastric muscle (DA), and the middle part of the sternocleidomastoid muscle (SCM) according to the SENIAM (Surface EMG for Non-Invasive Assessment of Muscles) guidelines [22 (link)]. Placing surface electrodes was performed by the same physiotherapist (author G.Z.). The reference electrode was placed on the forehead, in the center of the frontal bone. The arrangement of the electrodes symmetrically on the skin covering the examined muscles on both sides was preceded by palpation of the muscles during mandibular and head/neck movements. The electrodes on the superficial masseter muscle were located along the line from the mandible angle to the inferior border of the zygomatic bone. The electrodes on the anterior part of the temporal muscle were arranged along a perpendicular line from the superior border of the zygomatic bone to a cranial bone (in the projection of the sphenoid bone). The electrodes on the anterior bellies of the digastric muscle were placed approximately 1 cm medial to the base of the mandible. The electrodes on the sternocleidomastoid muscle were placed in the middle part of the muscle belly. The edges of the surface electrodes were in contact to maintain a constant spacing between the electrodes, as presented in Figure 1 [22 (link)].

In the laboratory, the length of the mandible (ML, ± 1 mm) was measured as a proxy of skeletal size (i.e. from the mesial border of the first incisor socket to the vertical part of the ramus, after having removed the flesh in these two points).
To estimate the age and tooth wear, mandibles were sectioned through a frontal plane between the entoconid-hypoconid and metaconid-protoconid of the first molar (M₁) using a circular diamond saw [26 ]. The age (in years) was estimated by counting the milky-coloured cement layers on the root pad of the sectioned mesial section of M₁, aided by a reflected light microscope at magnification x20 to x25 [27 ,28 ]. When M₁ was missing, or the cement layers poorly differentiated, M₂ was used and the age in years was estimated as the number of cement layers plus one.
Tooth wear was estimated by measuring, with the aid of a calliper and a magnifying glass, the thickness of the dentine on the sectioned mesial section of M₁ (molar height, MH, ± 0.1 mm) from the top of the cementum of the radicular pad to the middle point of the sectioned crown [7 ,26 ]. It has been noted that although crown formation in M₁ is fully complete at the age of 4 months [19 ,29 ], the completion of eruption and final positioning of the molar in the mandible does not take place until 3 years of age in red deer [19 ,20 (link)], and teeth also move in the mandible at very old age. Consequently, measuring molar height perpendicular from the mandible bone, labial or buccal, is not a reliable measurement of molar wear, especially in young age classes; by contrast, our MH measurement is independent of any movement of the molar in the mandible. We also measured crown size of the first permanent incisor (incisor height, IH, ± 0.1 mm), from the labial gingival sulcus to the median sagittal plane top of the crown [1 ].
Tooth wear was assessed as the negative relationship of crown height with age [7 ,26 ].

The performance for identification location of maxillofacial fractures in CT images was evaluated by precision, recall, F1 score, sensitivity, specificity and accuracy of the CNN-based multiclass image classification models. The accuracy performance for fracture detection of CNN-based object detection models was evaluated by precision, recall, F1 score, average precision (AP), and mean average precision (mAP). Receiver operating characteristic (ROC) and precision-recall curves were generated using a Python script. An ROC curve plotted by varying the operating threshold was used to assess the ability of the classification model in the discrimination of each class. An ROC curve provided the tradeoff between the sensitivity and 1-specificity. An area under the ROC curve (AUC) was used to summarize the diagnostic accuracy of each class. It was found to have good coverage accuracy over imbalance classes^{30 (link)}. The statistical analysis for multiclass image classification and object detection was calculated as follows^{31 (link)}: $$\text{Precision}=\frac{\text{TP}}{\text{TP}+\text{FP}}$$ $$\text{Recall}\left(\text{Sensitivity}\right)=\frac{\text{TP}}{\text{TP}+\text{FN}}$$ $$\text{F}1\text{score}=2\times \frac{\left(\text{Precision}\times \text{Recall}\right)}{\text{Precision}+\text{Recall}}$$ $$\text{Specificity}=\frac{\text{TN}}{\text{TN}+\text{FP}}$$ where the "frontal", "midface" and "mandible" fracture classes are positive classes and the “no Fx” class is a negative class.

True Positive (TP) is the number of "frontal", "midface", or "mandible" fracture classes that had a correct prediction or detection.

True Negative (TN) is the number of “no Fx” images that had a correct prediction or detection.

False Negative (FN) is the number of "frontal", "midface", or "mandible" fracture classes that had no prediction or detection.

False Positive (FP) is the number of “no Fx” images that had false prediction as fracture class images.

This study retrospectively analyzed the CT images data from 150 patients aged 18 years or older in the oral and maxillofacial clinic of regional trauma centers from 2016 to 2020 with a diagnosis of maxillofacial fractures. The inclusion criteria for collecting data from patients aged 18 years or older was due to the data availability in these centers. CT images confirmed maxillofacial fractures were based on a manual review of the clinical and radiological reports recorded in the trauma centers. The maxillofacial CT images were obtained with equipment from different manufacturers using standard imaging protocols. Two-dimensional planar reconstructions were performed in the frontal, sagittal, and oblique planes, parallel to the long axis of the orbits, hard palate and mandible. CT images contained a varied number of axial, sagittal and coronal views with slice thickness of 0.5–2 mm. in increments and a matrix size of 512 × 512 pixels. To develop the maxillofacial fracture detection models, CT images in this study were selected in a two-dimensional axial view of the bone window (window parameters – 2200/200 HU).
A total of 3,407 maxillofacial CT images of 150 patients admitted to trauma centers was divided into CT images containing maxillofacial fractures and CT images without fractures. Of these, 2407 CT images of maxillofacial fracture were distributed to three sites of the maxillofacial area: the frontal fracture of 712 images, the midfacial fracture of 949 images, and the mandibular fracture of 746 images. Another 1000 maxillofacial CT images without fractures were selected from slices of CT images without fracture lines or pathologic lesions.
All CT images were uploaded to the VisionMarker²² (Digital Storemesh, Bangkok, Thailand) server. VisionMarker is a private web application for image annotation; the public version is available on GitHub (GitHub, Inc., CA, USA). To build the CNN models, the maxillofacial fracture line or ground truth on CT images was reviewed and annotated by consensus of five oral and maxillofacial surgeons with more than 5 years of experience in maxillofacial trauma. For CT images containing maxillofacial fractures, rectangular bounding boxes were drawn around each fracture line and classified as frontal, midface and mandible class according to the location of fracture of frontal, midfacial and mandibular area, respectively (Fig. 1). And for CT images without fracture lines, all images were classified as no fracture. Manual annotations in 3 classes (frontal, midface and mandible) and no fracture (no Fx, without annotation) were used in the learning process of object detection, while multiclass classification (frontal, midface, mandible and no Fx classes) did not require bounding box annotations because the locations were not identified with a classifier. The bounded frontal, midface and mandibular fracture images were split by the accession number into the training, validation, and independent test sets using a 70:10:20 split, with a randomization by distribution to ensure an equal distribution of datasets.Figure 1

The CNN workflow of data set construction, model building and evaluation. CNN convolutional neural network, CT computed tomography, Fx fracture.

This retrospective study was carried out from the archives of the Oral and Maxillofacial Radiology section (from March 2017 to March 2021) at Manipal College of Dental Sciences, Manipal. We selected 1000 digital panoramic radiographs of individuals aged between 12 and 25 years. Radiographs of individuals belonging to the southern part of the state (South Indian population) were considered after verifying their address from medical records. The difference between the date of birth provided in the dental record and the date on which the radiograph was taken was considered to calculate the age of the individual. Radiographs with diagnostically acceptable images of intact mandibular second and third molars were included in the study. Radiographs with the third and second molars missing or obscured due to artifacts, trauma, or fracture lines of the mandible passing through these molars were excluded. The radiographs that showed various lesions, syndromes, and developmental disorders were also excluded.
The study was conducted after receiving approval from the Institutional Ethics Committee (I.E.C. No: 249/2021).