Table 3.

Advanced AI applications in oral surgery.

Authors	Summarized Abstract	Methods Used	Results	Conclusions
Tanikawa [32]	-AI systems predict facial morphology post orthognathic surgery and orthodontic treatment.	-Collected lateral cephalograms and 3-D facial images pre- and post-treatment to develop two AI systems (System S and System E) using landmark-based geometric morphometric methods combined with deep learning.	-Systems S and E had average errors of 0.94 mm and 0.69 mm. -Success rates for Systems S and E were 54% and 98%.	-Developed AI systems predict facial morphology after orthognathic surgery and orthodontic treatment. -AI systems are confirmed to be clinically acceptable for predicting facial morphology.
Jeong [34]	-CNNs judge soft tissue profiles for orthognathic surgery using facial photos.	-A comparative study with 822 subjects divided into two groups of 411 each -Front and side facial photographs were taken; the VGG19 CNN model was employed for analysis.	-CNNs achieved 89.3% accuracy in judging soft tissue profiles. -Precision, recall, and F1 scores were 0.912, 0.867, and 0.889.	-CNNs can accurately judge soft tissue profiles for orthognathic surgery. -Deep learning networks are valuable for screening in dental field.
Zhang [30]	-Artificial neural networks predict postoperative facial swelling after third molar extraction.	-Evaluated an artificial neural network’s accuracy in predicting postoperative facial swelling after mandibular third molar extraction. -Employed an improved conjugate gradient BP algorithm.	-Artificial neural network model predicted postoperative facial swelling with 98.00% accuracy.	-Improves conjugate gradient BP algorithm enhances prediction accuracy.
Kim [31]	-Investigated image patterns in cephalometric radiographs for orthognathic surgery diagnosis.	-Involved 960 patients split between those requiring orthognathic surgery (320) and non-surgical treatments (640). -Utilized CNN models ResNet-18, 34, 50, and 101.	-ResNet-18 had the best performance in screening with an AUC of 0.979. -Average success rates for ResNet models ranged from 91.13% to 93.80%.	-Study outlines characteristics needed for medical image-based decision-making models.
Choi [33]	-AI model determined mandibular third molar (M3) and inferior (alveolar nerve) IAN position in panoramic radiographs.	-Utilized 571 panoramic images to develop an AI model using ResNet-50 to determine the positional relationship between M3 and IAN.	-AI accuracy for true contact position was 72.32%, superior to OMFS. -AI accuracy for bucco-lingual position was 80.65%, surpassing OMFS specialists.	-AI model shows higher accuracy in bucco-lingual position determination. -AI can support clinicians in decision making for M3 treatment.