The initiation of radiologic analysis involves the selection of a high-quality and suitable imaging protocol. Imaging modalities including CBCT, CT, MRI, nuclear medicine, PET, and US 9. The quality of the image is crucial, as substandard or artifact-laden photographs might negatively impact the entire process.

After collecting image data, segmentation can be done manually, semi-automatically, or automatically. Image ROIs and VOIs are segregated. At this point, the target lesion, tumor, tissue, or anatomical structure is properly defined 10. Each segmentation approach has pros and cons. Most photo segmentation methods are manual or semi-automatic, including manual correction, although they have many limitations. Initially, manual segmentation is laborious. Manual segmentation time depends on target area size and data amount. Second, manual and semi-automated segmentation may have user-specific faults 11. Human mistake reduction is best with automatic segmentation. However, machine learning-based segmentation algorithms may fail if the training data lacks diversity. Clinical staff may be skeptical of automatic segmentation algorithms due to a lack of understanding of their operational mechanisms (the “black box” issue), requiring a post-procedure user verification. However, further study is needed to develop robust and generalizable automatic segmentation systems 12.

The quantitative analysis of segmented data calculates radiomics properties. Theory allows infinite probability computation of shape-based first, second, and higher order features 13. Shape-based examination includes ROI morphology, size, placement, vascularization, and spiculation. First-order statistics directly examine pixel or voxel intensity 14. Instead of intensity values, second-order features focus on spatial relationships between pixels or voxels. Understanding texture patterns and features is the goal 15. Wavelet, LoG, and LBP filters and transformations are used to extract higher-order features from the image, followed by image preprocessing. Re-extraction of first- or second-order characteristics reveals hidden patterns and complicated relationships 8.

Normalization and standardization are used to prepare the photos and extracted attributes for analysis, which ensures accuracy and reproducibility 2. In high-dimensional data sets, model correctness, consistency, and generalizability must be improved. Thus, photo noise and artifacts are identified. Radiological feature values at various scales greatly affect AI algorithm parameter stability. By synchronizing photos from different devices or protocols, findings are guaranteed to be comparable 16.

This large dataset should be augmented by a model that includes genomic profiles, histological characteristics, serum biomarkers, patient history, and diagnostic procedures to ensure data repeatability and accuracy. Only important elements are kept, while redundant, repetitive, or noise-inducing ones are removed. This level uses statistics and machine learning 17, 18.

Model development algorithms abound. The categorizing process requires a good model. Logistic regression, Support Vector Machine (SVM), random forests, and Extreme Gradient Boosting (XGBoost) distinguish benign and malignant tumors. Survival time is predicted using linear regression and survival trees. Convolutional neural networks (CNN) can handle large datasets. The literature seems to randomly select algorithms, while optimal practice involves substantial trial 16, 18, 19, 20.

Data is split into training and test sets before validation. The test set (20-30%) is smaller than the training set (70-80%). Validation is performed on clustered data at a given rate. For clinical approval, results must be validated using independent datasets, preferably from another institution 10, 20. Thus, independent external validation is the best model validation method. Small-scale pilot or early studies may not have independent validation data. These situations may require internal validation. In the literature, k-fold and leave one-out cross-validation (LOOCV) are the main internal validation methods 20.

The assessment of categorization performance is typically conducted using the area under the curve (AUC) 21. In datasets with class imbalance, AUC may not be a reliable performance indicator. A complete evaluation must include accuracy, sensitivity, specificity, Receiver Operating Characteristic (ROC) curve, and Matthews correlation coefficient 16. Following the assessment of classifications, regression and clinical validation analyses are performed and tested on real patient data for inclusion in the clinical decision support system 14.

The models must be tested using independent datasets from many centers or devices, regardless of previous procedures. The model must be generalizable on the training dataset and in clinical applications 22. Standardize the imaging parameters, segmentation methodologies, and feature extraction criteria 23. Before clinical use, the model's diagnostic precision, treatment efficacy, and prognosis prediction are evaluated. Finally, we must prospectively test the model in new patient cohorts and real-time clinical scenarios 24.

Panoramic radiographs usually detect jaw cysts and cancers. CBCT is the most common imaging modality in dentistry for thorough examination, however radiographic images alone cannot diagnose certain diseases due to their different clinical, radiographic, and histopathologic features. Indeed, CBCT is difficult for discriminating benign from malignant tumors. Numerous studies have developed deep learning models using 2D or 3D radiographic images to automatically identify various jaw cysts and malignancies in response to the need for accurate diagnosis 26, 27, 28, 29. Xu et al.30 developed artificial intelligence models to differentiate between odontogenic cysts and ameloblastomas by utilizing particular characteristics from CBCT images, achieving remarkably high accuracy rates. Sha et al. 31 developed a precise CBCT-based radiomics model to differentiate odontogenic cysts, keratocysts, and ameloblastomas. Additional imaging in this region may help maxillofacial radiologists identify jaw cysts and malignancies, improving early diagnosis and prognosis 25.

Salivary gland diseases are inflammatory, infectious, or neoplastic ailments that affect the parotid, submandibular-sublingual, and minor salivary glands. Salivary gland diseases are difficult to diagnose in dentistry and otolaryngology because they require clinical competence and diagnostic imaging (MRI, CT, US) 25. Radiomics analysis enables the automatic and objective differentiation of benign and malignant tumors by assessing lesion features, including tissue heterogeneity, form, and margins 32. Because MRI scans better depict soft tissues than other imaging modalities, deep learning models for salivary gland diseases have mostly used them 33. Since MRI is rarely used in dentistry and radiomics ultrasound research is limited, the models were mostly based on CT images 25. Yuan et al. 34 created a CNN model utilizing CT images to differentiate between pleomorphic adenoma and malignant parotid gland tumors; Zhang et al. 35 established a CNN model on CT images to distinguish between benign and malignant parotid gland tumors; Kise et al. 36 used a CNN model on CT scans to autonomously identify salivary gland fatty degeneration, a key indication of Sjögren's syndrome, with great accuracy.

Pulpal tissue is usually healthy but can be damaged by inflammation or trauma. This injury may migrate from the pulpal tissue to the periapical region, causing bone inflammation. Inflammatory periapical lesions alter alveolar bone internal imaging, indicating bone structure decrease, augmentation, or a combination. CBCT is often used in endodontics to investigate root canal blockages, dilations, contractions, perforations, canal morphology, and alveolar bone changes. Rosa et al. 37 used textural radiomic parameters from CBCT images to detect periapical granulomas and cysts with more accuracy than radiologic assessment. Gonzalez et al. 38 found textural and first-order radiomic differences in CBCT images for diagnosing periapical lesions based on volumetric size, cortical expansion, resorption, and shape.

The maxillary sinus, the biggest face sinus, is often involved in posterior tooth and dental implant procedures. A satisfactory treatment outcome requires accurate diagnosis and categorization of maxillary sinus disorders before sinus surgery. CT, CBCT, or MRI are used to detect mucoceles, sinusitis, mucus retention cysts, maxillary sinus cysts, and tumors and analyze their association with neighboring anatomical structures. Precise identification of maxillary sinus diseases on conventional imaging modalities can occasionally be challenging. Ito et al. 39 used CT to compare the tissue characteristics of maxillary sinus mucosal thickening in odontogenic and non-odontogenic maxillary sinusitis, finding significant differences. Shaujan et al. 40 employed CT scans to differentiate inverted papilloma from chronic rhinosinusitis polyps, and their radiomic models demonstrated high accuracy in this classification. Moreover, literature also presents CNN-based sinus segmentation 41 and morphological alterations in sinus mucosa 42.

TMJ disorders include issues with the joint itself (such as disc displacement, partial dislocation, and stiffness), conditions impacting the joint (like arthritis and specific cancers), fractures, developmental anomalies, and complications involving the masticatory muscles (including pain and tension). Inflammatory changes in the synovial membrane of the joints or masticatory muscles lead to variations in the joint complex caused by inflammatory mediators. This leads to degenerative and deformative changes in the cartilage matrix, subchondral bone, and related joint tissues. The diagnosis and treatment of TMJ disorders depend significantly on comprehensive clinical expertise, resulting in diverse viewpoints among doctors concerning diagnostic and therapeutic approaches. Recent research indicate that radiomics algorithms, especially those utilizing MRI and CBCT data, provide substantial insights in evaluating TMJ issues. Orhan et al. 43 categorized alterations in the condyle and disc dislocations using radiomicsing with MRI data and saw a significant concordance in their results. Ogawa et al. 44 found significant tissue changes linked with condylar flatness and erosion in CT and MRI images. The findings suggest that tissue analysis may help detect TMJ degeneration early.

Many clinical and radiographic indications are used to identify benign from malignant craniofacial diseases. However, some benign forms may mislead the doctor during the first diagnosis by seeming malignant on radiographs. Early-stage tumors may not show radiographic abnormalities. Therefore, radiomics can distinguish benign from malignant lesions by examining microscopic textural variations (heterogeneity, edge patterns, density distributions) that are imperceptible to the naked eye and converting them into mathematical data for artificial intelligence algorithms to analyze. Iqbal et al. 45 found that combining clinical exams, imaging modalities, and histopathologic tests improves oral lesion diagnoses and reduces misdiagnosis rates. Yuan et al. 46 developed an MRI-based radiomics algorithm and nomogram to predict preoperative outcomes in head and neck squamous cell carcinoma (SCC) patients. Research and technology predict that radiomics will reduce the need for intrusive therapies to distinguish benign from malignant illnesses, improve prognosis, and make diagnosis easier for physicians.

Although dental implants are a reliable treatment method for toothless areas, they can involve certain complications, as with any surgical procedure 47. These complications may occur during the surgical procedure or in the subsequent period. Failure may occur due to reasons such as incorrect application of the surgical intervention, insufficient bone quality, infection, or incorrect application of prosthetic rehabilitation 48. While monitoring success radiographically, physicians mostly use conventional radiographs (periapical and panoramic radiography) and CBCT images. However, this may not always provide sufficient results. The evaluation of dental implants using the radiomics method is an innovative approach that goes beyond traditional radiographic examinations and is based on obtaining quantitative data from medical images that cannot be distinguished by the human eye. This approach entails the examination of data acquired from three-dimensional imaging modalities, such as CBCT, with sophisticated mathematical algorithms 49. Radiomics analysis seeks to forecast the efficacy and enduring stability of the implant by delivering comprehensive insights into the tissue, density, and microarchitecture of the bone at the implant site. The integration of radiomics and machine learning techniques is regarded as a potential strategy for forecasting phenomena such as physiological bone remodeling (PBR) that may transpire shortly after implant implantation 50. This technology enables the prior identification of risk variables that may influence the effectiveness of implant treatment, allowing for the formulation of more precise tailored treatment programs.