Department of Nuclear Medicine, Kanazawa University Hospital, 13-1 Takara-machi, Kanazawa 920-8641, Japan.
EJNMMI Res. 2013 Dec 26;3(1):83. doi: 10.1186/2191-219X-3-83.
Artificial neural network (ANN)-based bone scan index (BSI), a marker of the amount of bone metastasis, has been shown to enhance diagnostic accuracy and reproducibility but is potentially affected by training databases. The aims of this study were to revise the software using a large number of Japanese databases and to validate its diagnostic accuracy compared with the original Swedish training database.
The BSI was calculated with EXINIbone (EB; EXINI Diagnostics) using the Swedish training database (n = 789). The software using Japanese training databases from a single institution (BONENAVI version 1, BN1, n = 904) and the revised version from nine institutions (version 2, BN2, n = 1,532) were compared. The diagnostic accuracy was validated with another 503 multi-center bone scans including patients with prostate (n = 207), breast (n = 166), and other cancer types. The ANN value (probability of abnormality) and BSI were calculated. Receiver operating characteristic (ROC) and net reclassification improvement (NRI) analyses were performed.
The ROC analysis based on the ANN value showed significant improvement from EB to BN1 and BN2. In men (n = 296), the area under the curve (AUC) was 0.877 for EB, 0.912 for BN1 (p = not significant (ns) vs. EB) and 0.934 for BN2 (p = 0.007 vs. EB). In women (n = 207), the AUC was 0.831 for EB, 0.910 for BN1 (p = 0.016 vs. EB), and 0.932 for BN2 (p < 0.0001 vs. EB). The optimum sensitivity and specificity based on BN2 was 90% and 84% for men and 93% and 85% for women. In patients with prostate cancer, the AUC was equally high with EB, BN1, and BN2 (0.939, 0.949, and 0.957, p = ns). In patients with breast cancer, the AUC was improved from EB (0.847) to BN1 (0.910, p = ns) and BN2 (0.924, p = 0.039). The NRI using ANN between EB and BN1 was 17.7% (p = 0.0042), and that between EB and BN2 was 29.6% (p < 0.0001). With respect to BSI, the NRI analysis showed downward reclassification with total NRI of 31.9% ( p < 0.0001).
In the software for calculating BSI, the multi-institutional database significantly improved identification of bone metastasis compared with the original database, indicating the importance of a sufficient number of training databases including various types of cancers.
基于人工神经网络（ANN）的骨扫描指数（BSI）是骨转移量的标志物，已被证明可以提高诊断准确性和重现性，但可能受到训练数据库的影响。本研究的目的是使用大量日本数据库修订软件，并与原始瑞典训练数据库进行比较，以验证其诊断准确性。
使用 EXINIbone（EB；EXINI Diagnostics）计算 BSI，使用瑞典训练数据库（n=789）。比较了来自单个机构的日本训练数据库的软件（BONENAVI 版本 1，BN1，n=904）和来自 9 个机构的修订版（版本 2，BN2，n=1,532）。使用来自另外 503 例多中心骨扫描的患者数据进行验证，包括前列腺癌（n=207）、乳腺癌（n=166）和其他癌症类型的患者。计算 ANN 值（异常概率）和 BSI。进行了接收器工作特征（ROC）和净重新分类改善（NRI）分析。
基于 ANN 值的 ROC 分析显示，从 EB 到 BN1 和 BN2 有显著改善。在男性（n=296）中，EB 的曲线下面积（AUC）为 0.877，BN1 为 0.912（p 值不显著（ns）与 EB 相比），BN2 为 0.934（p=0.007 与 EB 相比）。在女性（n=207）中，EB 的 AUC 为 0.831，BN1 为 0.910（p=0.016 与 EB 相比），BN2 为 0.932（p<0.0001 与 EB 相比）。基于 BN2 的最佳敏感性和特异性分别为男性 90%和 84%，女性 93%和 85%。在前列腺癌患者中，EB、BN1 和 BN2 的 AUC 同样高（0.939、0.949 和 0.957，p=ns）。在乳腺癌患者中，EB 的 AUC 得到改善（0.847），分别提高至 BN1（0.910，p=ns）和 BN2（0.924，p=0.039）。EB 与 BN1 之间的 ANN NRI 为 17.7%（p=0.0042），EB 与 BN2 之间的 NRI 为 29.6%（p<0.0001）。对于 BSI，NRI 分析显示总 NRI 为 31.9%（p<0.0001），向下重新分类。
在计算 BSI 的软件中，多机构数据库与原始数据库相比，显著提高了对骨转移的识别能力，表明包括各种类型癌症的足够数量的训练数据库的重要性。
