Research on Deep Learning-Based Intelligent Diagnosis and Treatment

This study is an imaging-based clinical artificial intelligence study that aims to develop, train, and validate deep learning models for craniomaxillofacial multi-modal imaging analysis and intelligent clinical decision support. Approximately 2,000 participants are planned to be enrolled. All participants will have craniomaxillofacial imaging data and related clinical information obtained during routine dental, orthodontic, oral and maxillofacial, or related clinical care.

The imaging data used in this study may include two-dimensional facial photographs, cone-beam computed tomography images, and three-dimensional facial surface scans. Related clinical data may include demographic information, clinical diagnosis, skeletal or dental classification, cephalometric measurements, treatment-related records, and available expert assessments. This study will focus on developing artificial intelligence models for craniomaxillofacial image classification, anatomical landmark detection, regional segmentation, abnormality recognition, and treatment-related decision support.

Before model development, all imaging data will undergo standardized preprocessing. Preprocessing procedures may include image de-identification, quality assessment, format conversion, image orientation correction, cropping, resolution standardization, grayscale or intensity normalization, spatial registration, and region-of-interest extraction. For two-dimensional images, standardized facial or radiographic regions will be defined according to the specific imaging task. For cone-beam computed tomography and three-dimensional facial surface data, preprocessing may include volumetric reconstruction, surface reconstruction, mesh processing, point-cloud processing, or anatomical structure extraction.

Reference labels will be derived from clinical medical records, clinician diagnosis, and expert manual annotation. For segmentation tasks, anatomical regions or structures will be manually annotated by trained clinicians or calibrated researchers. For landmark detection tasks, predefined anatomical landmarks will be annotated according to standardized craniomaxillofacial measurement protocols. For classification tasks, skeletal pattern, dental relationship, craniomaxillofacial morphology category, or other clinically meaningful classification labels will be determined based on expert assessment. Inter-observer and intra-observer consistency will also be evaluated to ensure the reliability of manual annotations and clinical labels.

The dataset of approximately 2,000 participants will be divided into training, validation, and testing datasets. Seventy percent of the data will be used for model training, 15% for model validation, and 15% for independent testing.

Deep learning models will include convolutional neural networks, attention-based models, and multi-task learning models. Model development will include hyperparameter optimization, regularization, data augmentation, and performance monitoring based on the validation dataset.

For image segmentation tasks, model performance will be evaluated using segmentation accuracy metrics, including the Dice similarity coefficient, Intersection over Union, pixel-wise accuracy, sensitivity, specificity, and, where applicable, boundary distance metrics such as Hausdorff distance or average surface distance. For image classification and diagnostic recognition tasks, model performance will be evaluated using accuracy, precision, recall, F1-score, sensitivity, specificity, positive predictive value, negative predictive value, receiver operating characteristic curve analysis, and area under the receiver operating characteristic curve. Confusion matrices will be used to describe the distribution of correct and incorrect classifications across diagnostic categories. For multi-class classification tasks, macro-averaged and weighted-averaged performance metrics may be reported. For anatomical landmark detection tasks, model performance will be evaluated using mean radial error, mean absolute error, successful detection rate within clinically meaningful distance thresholds, and agreement with expert annotations.

For treatment-related decision support tasks, model outputs will be compared with expert clinical judgment. Evaluation metrics will include accuracy, precision, recall, F1-score, AUC, and decision curve analysis.

Personnel involved in image annotation and data processing will receive relevant training before the start of the study. Access to identifiable personal information will be restricted to authorized research personnel. The analysis datasets used for model development will be de-identified or pseudonymized whenever possible.

The planned enrollment of approximately 2,000 participants is expected to support subgroup analyses and robustness testing when the distribution of imaging modalities and clinical categories is adequate. Missing, unavailable, uninterpretable, or inconsistent data will be recorded.

The artificial intelligence models developed in this study are intended to assist health care providers in craniomaxillofacial imaging analysis and clinical decision support. The models will not replace professional clinical diagnosis or individualized treatment planning by qualified clinicians. Any future clinical application of these models will require further validation, clinical review, and appropriate institutional or regulatory approval.