Shan Jie, Yang Yifei, Liu Hualian, Sun Zhaoyao, Chen Mingming, Zhu Zhichao
Resident, Department of Oral and Maxillofacial Surgery, The Third Affiliated Hospital of Soochow University, Changzhou, China.
Associate Chief Physician, Department of Oral and Maxillofacial Surgery, The Third Affiliated Hospital of Soochow University, Changzhou, China.
J Oral Maxillofac Surg. 2025 Feb;83(2):208-221. doi: 10.1016/j.joms.2024.10.018. Epub 2024 Oct 31.
Contrast-enhanced ultrasound (CEUS) is frequently used to distinguish benign parotid tumors (BPTs) from malignant parotid tumors (MPTs). Introducing machine learning may enable clinicians to preoperatively diagnose parotid tumors precisely.
We aimed to estimate the diagnostic capability of machine learning in differentiating BPTs from MPTs.
STUDY DESIGN, SETTING, AND SAMPLE: A retrospective cohort study was conducted at the Third Affiliated Hospital of Soochow University. Patients who underwent parotidectomy and CEUS for untreated parotid tumors were included. Patients with recurrent tumors, inadequate specimens, or chemoradiotherapy were excluded.
Predictor variable was preoperative diagnosis coded as BPTs and MPTs based on the support vector machine (SVM) algorithms, laboratory, and CEUS variables.
MAIN OUTCOME VARIABLE(S): Outcome variable was pathological diagnosis coded as BPTs and MPTs.
Covariate was demographics.
A senior surgeon labeled patients' tumors as BPTs or MPTs, creating a clinical diagnosis. Patients were randomly divided into training (70%) and testing (30%) sets. After developing the SVM models using the training set, we evaluated their diagnostic performance on the testing set with the area under the receiver-operating characteristic curve (AUC), accuracy, positive predictive value, negative predictive value, sensitivity, and specificity. Delong's test was used to compare the AUC of SVM models, laboratory, and CEUS variables.
The sample included 48 patients, and the testing set comprised 12 (25%) BPTs and 3 (6.25%) MPTs. Three CEUS variables (width, arrival time, and time to peak) and 3 laboratory variables (lymphocyte count, D-dimer, prognostic nutritional index) were identified through recursive feature elimination. Tested on the testing set, the SVM models with linear, polynomial, and radial kernels showed identical performance (AUC = 0.972, accuracy = 93.3%, positive predictive value = 75%, negative predictive value = 100%, sensitivity = 100%, specificity = 91.7%). They had larger AUC than SVM with sigmoid kernel (P = .18), width (P = .03), lymphocyte count (P = .02), D-dimer (P < .01), prognostic nutritional index (P = .03), arrival time (P = .02), time to peak (P = .04), CEUS diagnosis (P < .01), and clinical diagnosis (P < .01).
The SVM algorithm differentiated BPTs from MPTs better than laboratory and CEUS variables.
超声造影（CEUS）常用于区分腮腺良性肿瘤（BPTs）和腮腺恶性肿瘤（MPTs）。引入机器学习可能使临床医生能够在术前精确诊断腮腺肿瘤。
我们旨在评估机器学习在区分BPTs和MPTs方面的诊断能力。
研究设计、地点和样本：在苏州大学附属第三医院进行了一项回顾性队列研究。纳入了因未经治疗的腮腺肿瘤接受腮腺切除术和CEUS检查的患者。排除有复发性肿瘤、标本不足或接受过放化疗的患者。
预测变量是基于支持向量机（SVM）算法、实验室检查和CEUS变量编码为BPTs和MPTs的术前诊断。
结局变量是编码为BPTs和MPTs的病理诊断。
协变量是人口统计学资料。
一位资深外科医生将患者的肿瘤标记为BPTs或MPTs，形成临床诊断。患者被随机分为训练集（70%）和测试集（30%）。使用训练集建立SVM模型后，我们用受试者操作特征曲线下面积（AUC）、准确性、阳性预测值、阴性预测值、敏感性和特异性评估其在测试集上的诊断性能。使用德龙检验比较SVM模型、实验室检查和CEUS变量的AUC。
样本包括48例患者，测试集包括12例（25%）BPTs和3例（6.25%）MPTs。通过递归特征消除确定了3个CEUS变量（宽度、到达时间和峰值时间）和3个实验室变量（淋巴细胞计数、D-二聚体、预后营养指数）。在测试集上进行测试时，具有线性、多项式和径向核的SVM模型表现相同（AUC = 0.972，准确性 = 93.3%，阳性预测值 = 75%，阴性预测值 = 100%，敏感性 = 100%，特异性 = 91.7%）。它们的AUC大于具有Sigmoid核的SVM（P = 0.18）、宽度（P = 0.03）、淋巴细胞计数（P = 0.02）、D-二聚体（P < 0.01）、预后营养指数（P = 0.03）、到达时间（P = 0.02）、峰值时间（P = 0.04）、CEUS诊断（P < 0.01）和临床诊断（P < 0.01）。
SVM算法在区分BPTs和MPTs方面优于实验室检查和CEUS变量。
