超声图像融合专家知识的EfficientNet卵巢包块良恶性诊断方法

EfficientNet diagnosis method of benign and malignant ovarian masses based on ultrasound images incorporating expert knowledge

目的提出一种基于深度网络特征融合的分类方法,以提高良恶性分类的准确率,辅助医生提高术前诊断卵巢包块良恶性的准确率。方法纳入深圳市人民医院943幅经活检、手术病理等证实的患者术前卵巢超声图像,按照6∶2∶2的比例随机设置训练集、验证集和测试集。首先,提取医生勾画的感兴趣区域(region of interest,ROI)即包块图,用微调后的EfficientNet网络提取其深度特征;然后用基于Chan-Vese模型的水平集方法得到包块边缘轮廓图,再用微调后的EfficientNet网络提取其深度特征;接下来将包块图的深度特征和边缘轮廓图的深度特征分别归一化后并拼接为融合特征;最后,将融合特征输入到全连接层分类器中,将超声图像分为良恶性。结果本文提出的超声图融合专家知识的EfficientNet卵巢包块良恶性诊断方法在测试集上的准确度、特异度、敏感度和曲线下面积分别为0.81、0.78、0.88、0.91,全部优于当前主流的深度学习方法。结论该特征融合网络能够取得较好的分类效果,一定程度上能够为临床诊断卵巢包块的良恶性提供参考。

Objective To propose a classification method based on deep network feature fusion for improving the accuracy of benign and malignant classification, and to assist doctors improve the accuracy of preoperative diagnosis of benign and malignant ovarian masses.Methods The preoperative ovarian ultrasound images of 943 patients confirmed by biopsy and surgical pathology in Shenzhen People’s Hospital are included,and the training set, validation set and test set are randomly set according to the ratio of 6 ∶ 2 ∶ 2. First, extractthe region of interest(ROI) delineated by the doctor, that is, the block map, and use the fine-tuned EfficientNet network to extract the depth features of the block map;then use the level set segmentation method based on the Chan-Vese model to obtain the block. Edge contour map, and then use the fine-tuned EfficientNet network to extract its depth features;next, the depth features of the block map and the edge contour map are respectively normalized and spliced into fusion features;finally,the fusion features are input into the fully connected layer classifier, ultrasound images are classified into benign and malignant.Results The accuracy, specificity, sensitivity and area under curve of the EfficientNet method for the diagnosis of benign and malignant ovarian masses based on the fusion of expert knowledge of ultrasound images proposed in this paper were 0.81,0 on the test set, respectively 0.78, 0.88, 0.91, all of which are better than the current mainstream deep learning methods.Conclusions The network can achieve good classification effect, and to a certain extent can provide a reference for clinical diagnosis of benign or malignant ovarian masses.