基于深度学习的核素骨显像中骨转移瘤的诊断及评估

Automatic diagnosis and assessment of bone metastases on bone scans based on deep learning

目的:基于深度学习对^(99)Tc^(m)-亚甲基二膦酸盐(MDP)全身骨显像图像中骨转移瘤进行智能诊断,并设计肿瘤负荷的定量评估指标。方法:回顾性纳入同济大学附属第十人民医院核医学科2018年3月至2019年7月间621例患者(男389例、女232例;年龄12~93岁)的骨显像图像,分为骨转移瘤组和非骨转移瘤组。从2组分别抽取80%作为训练集,余20%作为测试集。利用深度残差卷积神经网络ResNet34构建骨转移瘤诊断分类及分割模型。计算灵敏度、特异性、准确性以评估分类模型性能,分析分类模型在<50岁(15例)、≥50且<70岁(75例)及≥70岁(33例)组的性能差异。利用模型分割骨转移瘤区域,以骰子系数评估分割模型结果与人工标注结果的比对。计算骨显像肿瘤负荷系数(BSTBI)以定量评估骨转移瘤肿瘤负荷。结果:骨转移瘤图像280例,非骨转移瘤图像341例;其中,训练集498例,测试集123例。诊断分类模型识别骨转移瘤的灵敏度、特异性及准确性分别为92.59%(50/54)、85.51%(59/69)和88.62%(109/123)。分类模型在<50岁组表现最佳(灵敏度2/2,特异性12/13,准确性14/15),其特异性在≥70岁组中最低(8/12)。分割模型中,骨转移瘤区域骰子系数为0.739,膀胱区域骰子系数为0.925,模型在3个年龄组表现相当。初步结果显示,BSTBI随病灶数目的增多、^(99)Tc^(m)-MDP摄取程度的增高而增大。构建的骨转移瘤智能诊断模型从输入原始数据到最终完成BSTBI计算所需时间为(0.48±0.07)s。结论:基于深度学习的骨转移瘤智能诊断模型能较准确地识别骨转移瘤、进行自动区域分割及计算肿瘤负荷,为骨显像图像的解读提供了新方法。研究提出的BSTBI有望成为骨转移瘤肿瘤负荷的定量评估指标。

Objective To develop an approach for the automatic diagnosis of bone metastasis and to design a parameter of quantitative evaluation for tumor burden on bone scans based on deep learning technology.Methods A total of 621 cases(389 males,232 females,age:12-93 years)of bone scan images from the Department of Nuclear Medicine in Tenth People′s Hospital of Tongji University from March 2018 to July 2019 were retrospectively analyzed.Images were divided into bone metastasis group and non-bone metastasis group.Eighty percent of the cases were randomly extracted from both groups as the training set,and the rest of cases were used as the test set.A deep residual convolutional neural network ResNet34 was used to construct the classification model and the segmentation model.The sensitivity,specificity and accuracy were calculated and the performance differences of the classification model in different age groups(15 cases of<50 years,75 cases of≥50 and<70 years,33 cases of≥70 years)were analyzed.The regions of metastatic bone lesions were automatically segmented by the segmentation model.The Dice coefficient was used to evaluate the effect of the segmentation model and the manual labeled results.Finally,the bone scans tumor burden index(BSTBI)was calculated to assess the tumor burden of bone metastases.Results There were 280 cases with bone metastases and 341 cases with non-bone metastases,including 498 in training set and 123 in test set.The classification model could accurately identify bone metastases,with the sensitivity,specificity and accuracy of 92.59%(50/54),85.51%(59/69)and 88.62%(109/123),respectively,and it performed best in the<50 years group(sensitivity,2/2;specificity,12/13;accuracy,14/15).The specificity in the≥70 years group(8/12)was the lowest.The Dice coefficient of bone metastatic area and bladder area were 0.739 and 0.925 in the segmentation model,which performed similarly in the three age groups.Preliminary results showed that the value of BSTBI increased with the increase of the number of bone metastatic lesions and the degree of^(99)Tc^(m)-MDP uptake.The machine learning model in this study took(0.48±0.07)s for the entire analysis process from input to the final BSTBI calculation.Conclusions The deep learning based on automatic diagnosis framework for bone metastases can automatically and accurately identify segment bone metastases and calculate tumor burden.It provides a new way for the interpretation of bone scans.The proposed BSTBI may be used as a quantitative evaluation indicator in the future to assess the tumor burden of bone metastases based on bone scans.