增强CT图像的深度学习分类技术在卵巢肿瘤良恶性预测中的应用

Application of deep learning classification technology of enhanced CT images in the prediction of benign and malignant ovarian tumours

目的 探讨利用患者术前卵巢肿瘤的增强CT图像结合深度学习技术预测卵巢包块良恶性预测的可行性。方法 收集2018年7月至2022年5月在四川省人民医院因卵巢肿瘤入院拟行手术治疗的患者,获取术前盆腹腔增强CT图像分为训练集和测试集。训练集:恶性患者50例1698幅图像,良性患者26例1293幅图像。测试集:恶性患者25例1284幅图像,良性患者13例766幅图像。由影像专科主治医师标注训练集标签,输入模型进行训练,测试集标签则用本研究同时建立的改良的U-net模型自动分割。训练完成后,分别由影像专科主治医师与DL模型分别对测试集进行良恶性诊断,对比两者的预测效果。结果 DL模型在分类测试中将25例恶性肿瘤病例全部判定为恶性,13例良性肿瘤有3例被误判为恶性,模型预测的灵敏度为100%,特异度为76.9%,准确度为92.1%,AUC为0.88。影像科主治医师在测试中将25例恶性肿瘤24例判定为恶性,13例良性肿瘤有2例误判为恶性,灵敏度为96%,特异性为84.6%,准确性为92.1%。DL模型预测效果同影像专科医师比较,差异无统计学意义(P>0.05),Kappa值>0.75,两者诊断效果有高度一致性。结论 基于增强CT图像的深度学习分类技术用于卵巢肿瘤术前的良恶性诊断是可行的。

Objective To explore the feasibility of using enhanced CT images of patients with preoperative ovarian tumors combined with deep learning(DL) technology to predict the benign and malignant of ovarian masses.Methods The patients hospitalized in our hospital for ovarian tumors from July 2018 to May 2022 were collected. The preoperative pelvic and abdominal enhanced CT images were obtained. The images were divided into training set and test set. The training set included 1698 images of 50 malignant patients and 1293 images of 26 benign patients. The test set included 1284 images of 25 malignant patients and 766 images of 13 benign patients. The training set label was labeled by the attending physician of the imaging specialty, and the input model was trained. The test set label was automatically segmented with the improved U-net model established at the same time in this study. After training, the attending physician of imaging specialty and DL model were respectively used to diagnose the benign and malignant of the test set, and the predictive effects of the two were compared.Results In the classification test of DL model, 25 cases of malignant tumors were all judged as malignant and 3 cases of 13 benign tumors were misjudged as malignant. The sensitivity, specificity, accuracy and area under curve(AUC) of ROC analysis for the model were 100%, 76.9%, 92.1% and 0.88 respectively. In the test, the chief physician of the imaging department judged 24 of 25 malignant tumors as malignant, and 2 of 13 benign tumors as malignant. The sensitivity was 96%, specificity was 84.6%, and accuracy was 92.1%. There was no significant difference in predictive effect between imaging specialists and DL model(P > 0.05). The kappa value was > 0.75. Thus, the diagnostic effect of the two models was highly consistent.Conclusions The deep learning classification technique of enhanced CT images is feasible for the preoperative diagnosis of benign and malignant ovarian tumors.