Preoperative Differentiation of Jaw Cystic Lesions Based on Radiomics From Computed Tomography Images: A Multicenter, Prospective Machine Learning Study

Jawbone cystic lesions include odontogenic tumors and non-tumorous cystic lesions occurring within the jawbone, with ameloblastoma being the most common among the former, and odontogenic and non-odontogenic cysts among the latter. Currently, the treatment focus varies for different types of jawbone cystic lesions. Ameloblastomas, which may recur and metastasize, are primarily treated with surgical excision, while cystic lesions are more broadly treated with procedures like curettage and marsupialization. Therefore, accurate preoperative differential diagnosis of various jawbone lesions and the subsequent selection of appropriate treatment plans are crucial for achieving optimal patient outcomes. Inappropriate treatment choices may delay the condition or lead to overtreatment, affecting the patient's quality of life. At present, there is still a lack of an objective and accurate standard and differential diagnosis protocol for the treatment of jawbone cystic lesions, making the establishment of an objective and scientific preoperative diagnostic prediction model of significant clinical importance. In previous research, we successfully developed an effective predictive diagnostic model by integrating machine learning techniques with computed tomography (CT) radiomics, achieving a maximum AUC value >0.8, indicating good predictive performance and clinical reference value. In the current study, we aim to conduct a multicenter, prospective machine learning study to further enhance the model's predictive diagnostic performance and assist clinical diagnosis and treatment.

This study is designed as a multicenter, prospective machine learning study, involving 300 patients with jawbone cystic lesions across 12 centers, as detailed in the list of collaborating institutions. Based on our previous investigation of the actual diagnostic and treatment conditions at each research center, we plan to utilize different types of imaging data for grouping according to the imaging examinations conducted, and to standardize the processing of imaging data from different units and types for subsequent work. Sun Yat-sen Memorial Hospital of Sun Yat-sen University will serve as the main center, with other institutions as sub-centers. The specific grouping is as follows: the spiral CT group includes six general hospitals; the cone beam CT (CBCT) group includes one general hospital and five specialized dental hospitals.

During the study, after enrolling patients who meet the inclusion criteria, we will collect maxillofacial CT imaging data, import them into the software (LIFEx version 6.30), and delineate the region of interest (ROI). Radiomic features within the ROI will be extracted using Pyradiomics software, selected, and used for preoperative diagnostic predictions with the existing model. After surgical treatment, the pathological results of the lesions will be tracked and recorded. If conditions permit, the model's predictive performance can be further optimized in phases during the study, or methodological adjustments and reconstructions of the predictive model can be attempted using all available data to achieve a more ideal preoperative diagnostic prediction.