基于深度学习和图像融合诊断甲状腺结节良恶性的研究

Differential diagnosis between benign and malignant thyroid nodules based on deep learning and image fusion

目的设计一种基于甲状腺超声图像深度学习的人工智能模型,对比设计的模型同目前主流人工智能诊断模型、超声科专家诊断甲状腺结节的价值。方法收集行手术治疗的2421例甲状腺肿瘤患者的术前超声图像共3493张;对图片进行预处理后,随机选取2591张图像通过重采样和图像融合技术作为人工智能模型训练集,288张图像作为验证集,614张图像作为测试集。以术后病理结果为金标准,统计分析主流的DenseNet、VGGNet、ResNet、改进的ResNet模型、超声科专家诊断甲状腺结节良恶性的准确性、敏感度、特异度、受试者工作特征(ROC)曲线下面积(AUC)。结果DenseNet模型诊断甲状腺结节准确性为73.1%、敏感度为61%、特异度为81%、阳性预测值0.79、阴性预测值0.67,AUC为0.79;VGGNet模型诊断甲状腺结节准确性为73.6%、敏感度为67%、特异度为78%、阳性预测值0.78、阴性预测值0.65,AUC为0.80;ResNet模型诊断甲状腺结节准确性为72.3%、敏感度为68%、特异度为75%、阳性预测值0.77、阴性预测值0.63,AUC为0.81;改进的ResNet模型组诊断甲状腺结节准确性为74.4%、敏感度为65%、特异度为81%、阳性预测值0.80、阴性预测值0.68,AUC为0.82。超声科专家诊断甲状腺结节准确性为84%、敏感度为78%、特异度为90%、阳性预测值0.90、阴性预测值0.78,AUC为0.89。结论通过重采样和图像融合技术,研究设计的模型对甲状腺结节的诊断效能高于基线模型,且优于主流模型,对比超声科专家诊断效能仍有不足,需要持续改进。

Objective To design an artificial intelligence model based on deep learning of thyroid ultrasound images,and to compare the designed model with the current mainstream artificial intelligence diagnosis model and the value of ultrasound experts in diagnosing thyroid nodules.Methods This study collected 3,493 preoperative ultrasound images of 2,421 patients with thyroid tumors who received surgical treatment.After preprocessing the pictures,2,591 images were randomly selected through resampling and images fusion technology as the artificial intelligence model training set;288 images were used as the verification set,and 614 images as the test set.Taking the postoperative pathological results as the gold standard,the accuracies,sensitivities,specificities,ROC curve areas of mainstream DenseNet,VGGNet,ResNet,improved ResNet model,and ultrasound experts opinion in the diagnosis of benign and malignant thyroid nodules were analyzed.Results The accuracy of the DenseNet model for diagnosing thyroid nodules was 73.1%,the sensitivity was 61%,the specificity was 81%,the positive predictive value was 0.79,the negative predictive value was 0.67,and the area under the ROC curve was 0.79.The accuracy of the VGGNet model for diagnosing thyroid nodules was 73.6%,the sensitivity was 67%,the specificity was 78%,the positive predictive value was 0.78,the negative predictive value was 0.65,and the area under the ROC curve was 0.80.The accuracy of the ResNet model for diagnosing thyroid nodules was 72.3%,the sensitivity was 68%,the specificity was 75%,the positive predictive value was 0.77,the negative predictive value was 0.63,and the area under the ROC curve was 0.81.The accuracy of the improved ResNet for diagnosing thyroid nodules was 74.4%,the sensitivity was 65%,the specificity was 81%,the positive predictive value was 0.80,the negative predictive value was 0.68,and the area under the ROC curve was 0.82.The accuracy of ultrasound specialists for diagnosing thyroid nodules was 84%,the sensitivity was 78%,the specificity was 90%,the positive predictive value was 0.90,negative predictive value was 0.78,and the area under the ROC curve was 0.89.Conclusion Through resampling and image fusion technology,the diagnostic efficiency of the research designed model for thyroid nodules is higher than the baseline model,and better than the mainstream models.Compared with senior sonographers,the diagnostic efficiency is still insufficient and needs improvement.