Background
Bone scintigraphy has been accepted as a means to identify bone metastases associated with various types of cancer. Even after the advent of single-photon emission computed tomography combined with X-ray computed tomography, whole-body bone imaging is a standard method to survey the existence and extent of bone metastasis. Moreover, although bone scan interpretation may be performed on visual interpretation of whole-body images, an appropriate quantitative approach has been expected. While initial detection of bone metastases is important, quantification of progress of metastasis that results in patients’ disability, pain, pathological fractures, and mortality would be also beneficial [
1,
2]. However, there had been no definite imaging method that reflected metastatic disease burden and treatment effect before the advent of bone scan index (BSI) proposed at Memorial Sloan-Kettering Cancer Center [
3].
The BSI was developed as a marker of the spread for bone metastasis, which is a fraction of bones involved by a tumor and which realizes the regional distribution of the lesions [
4]. The software program for calculating BSI using the neural network system has also been developed using whole-body images with a Swedish database [
5]. They successfully applied automatic segmentation of the skeletal regions and automatic detection and feature extraction of hot spots using the neural network system. However, the diagnostic accuracy is potentially influenced by training databases. Whether the same database can be used universally in any study population is yet to be determined. The initial version using a Japanese database showed promising results with a revised database, but it was based on a single-center database [
6,
7].
The aims of this study were to create a multi-center Japanese database based on a large number of subjects with and without definite bone metastasis and to test the diagnostic accuracy compared with the original European database. In addition, to understand the characteristics of diagnostic accuracy based on the new database, a net reclassification improvement analysis was performed [
8].
Discussion
This study was performed as a multi-center project to establish a software program by incorporating a database that includes large number of patients with bone metastasis from various cancer types. While the software based on a Japanese single-center database improved the diagnostic accuracy compared with the software based on the original European database, the multi-center database including 1,532 patients further enhanced the diagnostic accuracy. The large training database also made it possible to use gender-specific analysis in BN2.
In addition to the diagnostic use of the software, BSI provides a quantitative measure that reflects the tumor burden expressed as a percentage of total body skeletal mass. The initial study started at Memorial Sloan-Kettering Cancer Center in patients with prostate cancer and showed good reproducibility and a parallel change with prostate-specific antigen [
3,
4]. BSI has been proved to contain prognostic information in addition to that of conventional prognostic markers such as clinical T stage, Gleason score, and prostate-specific antigen, and it has therefore drawn the attention of oncologists and urologists [
10]. When prostate cancer patients were stratified into, for example, high, intermediate, and low BSI groups, significant differences in survival rate were demonstrated [
11‐
13]. On-treatment changes in BSI could be a good response indicator rather than prostate-specific antigen alone in patients with castration-resistant metastatic prostate cancer.
The quantification of bone scans became practical by the use of a computer-aided diagnosis system with ANN, since the quality of visual bone scan interpretation varied according to readers’ experiences [
14]. When the segmentation of the skeletons, hot spot detection, evaluation of the characteristics of hot areas, and summed quantitative indexes were available with an automatic method, the reproducibility could be enhanced [
5,
15]. In a study using EB, a close correlation was demonstrated between manual and automated BSI measurements, and the merit of the latter was 100% reproducibility [
9]. Owing to simple application, BSI could be incorporated into clinical practice, while patients were diagnosed, treated, and followed up on.
Training databases are essential for a neural network system to diagnose bone metastases. In this study, we used only Japanese patients with definite diagnosis for the existence of bone metastasis. In addition to patient-based diagnostic accuracy, all the hot areas were confirmed by other imaging modalities and/or follow-up bone scans. Since the BN1 included only 141 (16%) patients with bone metastases from one hospital, it was increased to 638 (42%) patients from nine hospitals. The number of hot spots in ribs, for example, was increased from 2,303 (metastasis 50%) in BN1 to 3,294 (metastasis 43%) in BN2, which contributed to enhancing the learning volumes. When it is utilized in a number of hospitals, the multi-center database judged by multiple experts would be beneficial for enhancing diagnostic accuracy in computer learning.
The larger collection of databases including various cancer types is essential for obtaining appropriate BSI values. When we used the EB on Japanese patients for the first time, hot spots indicating high probability of abnormality were frequently noticed in the skull, shoulder joints, and lumbar vertebrae. These regions included diffuse metabolic accumulation in the skull of female patients and degenerative changes in the vertebrae and joints. About half (n = 425) of the Swedish database was from prostate cancer and 28% (n = 217) from breast cancer. In contrast, the Japanese databases for BN2 included 29% (n = 451) from prostate cancer, 41% (n = 624) from breast cancer, and 30% (n = 457) from other cancer types. The BN2 databases, therefore, included various cancer types and were closer to the usual clinical environment. From the viewpoint of Japanese population-specific databases, not only the physical stature but also the incidence of degenerative or deformative bone changes might differ between Swedish and Japanese subjects. When EB and BN1 were compared, NRI analysis with ANN showed that BN1 increased negative cases in patients without metastasis, indicating significantly decreased false-positive cases. BN2 further adjusted the diagnostic accuracy and reclassified the metastatic lesions into the higher ANN groups. With respect to the influence of revisions on BSI, the NRI analysis showed that reclassification was downward in both metastatic and non-metastatic groups. However, reclassification of non-metastatic patients into the lower risk BSI seemed to have meaning, and total net reclassification was improved in one third of the patients. The final effect of revision on predicting prognosis should be confirmed in future follow-up studies.
Notable effects of training databases differ among prostate, breast, and other cancers. The differences among cancer types seemed to be related to osteoblastic and osteolytic activity of the bone metastases and their imbalance in regulation [
16]. Quantitative measurement of bone metastasis or BSI has most widely been used in patients with prostate cancer [
9,
10,
12,
17,
18]. Prostate cancer shows typical osteoblastic metastasis based on radiological findings, though it is also associated with osteoclastic process and bone resorption. The bone scan appearance in prostate cancer reveals multiple hot spots and even the so-called superscan in extreme situations. Detecting all metastatic hot areas is important when demanding an overview of the whole amount of metastasis in prostate cancer. Therefore, the diagnostic accuracy in identifying bone metastasis was high even with EB, and further improvement by BN1 and BN2 was not achieved. In contrast, breast cancer commonly metastasizes to bones and destroys its structure, which causes both osteolytic and osteoblastic appearance in bones. The bone scan might show relatively mild activity or even cold areas in the pure osteolytic lesions. Higher fractions of breast and other cancer types in BN2 as compared with EB were also noted, namely, non-prostate cancer, 46% (
n = 364) for EB, 70% (
n = 637) for BN1, and 71% (
n = 1081) for BN2 (Table
1). To enhance the diagnostic accuracy in breast cancer metastasis, decreasing false-positive hot spots had practically important meaning, and it explained why the diagnostic improvement was obtained in BN1 and BN2 as compared with EB.
Limitations
The detection of metastasis was based on the hot areas, and cold lesions were not included for training the ANN system. However, since most of the diagnosis of the bone metastasis was made by the accumulation of 99mTc-MDP, the utility of BSI would not be substantially changed. Although database training was performed using all subjects, specific cancer type-based training, for example, prostate cancer-specific and breast cancer-specific training databases, could be applied. This process requires considerable time for separate training and will be studied in future works. Finally, when the NRI analysis is performed based on skeletal-related events, instead of diagnosis of metastases, true values of BSI will be confirmed in the future.
Even when we consider the possibility of 18 F NaF positron emission computed tomography in the future, a similar approach using ANN and new training databases might be an interesting project. What kinds of algorithm of ANN system are appropriate for tomographic images and/or maximum intensity projection images should be investigated.
Acknowledgements
This work was performed as a multi-center study for creating a new database involving nine institutions and supported by the cooperation of medical doctors and technologists in all institutions. Within the multi-center study, the authors would like to express their gratitude for the cooperation of Dr. Tsuyoshi Kawano (Kanagawa Cancer Center); Dr. Kiichiro Kodaira (International Medical Centre-Comprehensive Cancer Centre, Saitama Medical University); Takeshi Ono, RT (Shikoku Cancer Center); Dr. Koichiro Tsugawa, Dr. Istuko Okuda, Dr. Yasuyuki Kojima, and Yoshiaki Maehara, RT (St. Marianna University School of Medicine); Dr. Kiyoshi Koizumi, Dr. Tsuyoshi Hashimoto, Dr. Kunihito Suzuki, and Kenji Uchida, RT (Tokyo Medical University); Dr. Osamu Manabe (Hokkaido University); and Dr. Kazuhiro Nagao, Dr. Osamu Tokuda, and Yona Oishi, RT (Yamaguchi University). The authors wish to thank Mr. Mattias Ohlsson and Mr. Jens Richter (EXINI Diagnostics, Lund, Sweden), who were involved in software development and training with databases. The authors also thank Mr. Kazunori Kawakami and Mr. Akihiro Kikuchi (FUJIFILM RI Pharma, Co. Ltd, Tokyo, Japan) for their assistance in accumulating all the databases and for their valuable suggestions. They also thank Mr. Ronald Belisle for his editorial assistance and preparation of the manuscript.
Competing interests
LE is employed by and a shareholder of EXINI Diagnostics AB. KN has a collaborative research with FUJIFILM RI Pharma, Co. Ltd, Japan. While FUJIFILM RI Pharma is involved in the distribution of BONENAVI software in Japan, KN has no relationship relevant to the contents of this study to disclose.
Authors’ contributions
KN drafted the manuscript and LE and HH edited it. All the authors participated in the study design, and all but LE were involved in the final diagnosis and interpretation of bone scans. KN performed the statistical analysis of the data and YN confirmed it. All authors read and approved the final manuscript.