基于卷积神经网络的涎腺肿瘤超声图像分类研究

Research on ultrasonic image classification of salivary gland tumors based on convolutional neural network

为实现更准确地鉴别诊断涎腺肿瘤良恶性,避免不必要的穿刺或活检,本研究提出一种基于卷积神经网络(CNN)的涎腺肿瘤常规超声图像分类方法,并结合其超声图像特征,提高了鉴别涎腺肿瘤良恶性的准确率。将984张涎腺肿瘤超声图像分为训练集689张、验证集197张、测试集98张,结果显示训练集中该分类方法的最高准确率为92.43%,测试集中该分类方法鉴别诊断涎腺肿瘤良恶性的曲线下面积、准确率、灵敏度、特异度、阳性预测值、阴性预测值分别为0.863、85.44%、86.67%、86.27%、0.701、0.915,显示出对恶性样本具有较好的学习效果,证实了该模型鉴别诊断涎腺肿瘤良恶性的可行性,以及通过深度学习与人工提取特征图像结合的方法可以获得更高的识别准确率。本研究提出的方法可较高效、准确地对涎腺肿瘤良恶性进行分类,使目标病灶的检出及鉴别诊断更加直接、清晰且客观,辅助提高使用者诊断效率。

In order to achieve a more accurate differential diagnosis of benign and malignant salivary gland tumors by ultrasound and to avoid unnecessary puncture or biopsy,this study proposed a conventional ultrasound image classification method of salivary gland tumors based on a convolutional neural network(CNN),which combined with the features of salivary gland tumor ultrasound images to improve the accuracy of the differential diagnosis of benign and malignant salivary gland tumors.A total of 984 ultrasound images of salivary gland tumors were divided into training sets(n=698),validation sets(n=197)and test sets(n=98).The results showed that the highest accuracy of the classification method in the training set was 92.43%.The area under the curve,accuracy,sensitivity,specificity,positive predictive value and negative predictive value of the classification method in the test set were 0.863,85.44%,86.67%,86.27%,0.701 and 0.915,respectively.It shows a good learning effect on malignant samples.The results confirm that the feasibility of this model in the differential diagnosis of benign and malignant salivary gland tumors was verified,as well as the higher recognition accuracy that can be obtained by combining deep learning with manual extraction of feature images.The method proposed in this study can classify benign and malignant salivary gland tumors more efficiently and accurately,making the detection and differential diagnosis of target lesions more direct,clear and objective,and assisting in improving the diagnostic efficiency of users.