Related Experiment Video
Updated: May 4, 2026

Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Enhanced diagnostic accuracy for quantitative bone scan using an artificial neural network system: a Japanese
Kenichi Nakajima1, Yasuo Nakajima, Hiroyuki Horikoshi
1Department of Nuclear Medicine, Kanazawa University Hospital, 13-1 Takara-machi, Kanazawa 920-8641, Japan. nakajima@med.kanazawa-u.ac.jp.
This study updated computer software used to detect bone cancer spread. By training the system with a larger, more diverse set of Japanese patient data, the researchers improved the accuracy of identifying bone metastases compared to the original Swedish-based software. The new version performed better across different patient groups, including those with breast and prostate cancer.
Area of Science:
- Diagnostic imaging and artificial neural network research within oncology
- Medical informatics and bone scan index methodology
Background:
No prior work had resolved how regional training data influences automated bone scan interpretation. Prior research has shown that computational tools can quantify metastatic disease burden. That uncertainty drove the need for diverse training sets. It was already known that initial software models relied heavily on specific European patient populations. This gap motivated an investigation into whether local databases improve diagnostic performance. Prior studies suggested that software performance might vary across different geographic regions. No consensus existed regarding the optimal size of training datasets for these algorithms. This study addresses the limitations of existing diagnostic tools by incorporating broader clinical data.
Purpose Of The Study:
The aim of this study was to revise diagnostic software using a large number of Japanese databases. Researchers sought to address potential biases inherent in software trained on limited, non-local patient populations. The project specifically focused on improving the bone scan index, a marker for metastatic disease. By expanding the training set, the team intended to enhance the reproducibility of automated diagnostic results. The study also sought to validate the revised software against the original Swedish training database. This comparison was designed to highlight the impact of database size and diversity on diagnostic performance. The researchers aimed to provide a more accurate tool for identifying bone metastasis across various cancer types. This effort was motivated by the need for more robust, regionally adapted diagnostic systems in clinical oncology.
Main Methods:
The review approach involved comparing three distinct software versions for calculating metastatic burden. Researchers utilized the original Swedish training database as the baseline for performance evaluation. They then integrated a single-institution Japanese database to create the first updated version. A subsequent revision incorporated data from nine different medical centers to enhance model robustness. The team validated these tools using a separate cohort of over five hundred multi-center scans. Statistical evaluation relied on receiver operating characteristic curves to determine diagnostic accuracy. Net reclassification improvement analysis provided a quantitative measure of performance gains between versions. This systematic comparison allowed for a precise assessment of how database diversity influences automated diagnostic outcomes.
Main Results:
Key findings from the literature indicate that the multi-center database significantly improved diagnostic accuracy compared to the original model. The area under the curve for the final software version reached 0.934 in men and 0.932 in women. These values represent a statistically significant improvement over the baseline system. In breast cancer patients, the diagnostic performance increased from an area under the curve of 0.847 to 0.924. The net reclassification improvement analysis showed a total gain of 29.6 percent when comparing the final version to the original. The researchers achieved an optimum sensitivity of 90 percent and specificity of 84 percent for male patients. For female patients, the system reached 93 percent sensitivity and 85 percent specificity. These results confirm that larger, more diverse training sets lead to more reliable automated bone scan interpretations.
Conclusions:
The researchers propose that multi-institutional training data significantly enhances the identification of bone metastases. This study demonstrates that updating software with diverse patient records improves diagnostic precision. The authors suggest that the revised version outperforms the original model across both male and female cohorts. The findings indicate that incorporating various cancer types is necessary for robust performance. The authors conclude that the updated software provides superior classification compared to the previous iteration. This work highlights the necessity of large-scale databases for training reliable diagnostic systems. The authors report that the multi-center approach leads to better clinical outcomes in bone scan interpretation. These results support the continued development of regionally adapted diagnostic software for oncology.
Frequently Asked Questions
The researchers propose that the updated software, BN2, significantly improves diagnostic accuracy by utilizing a larger, multi-institutional Japanese training database. This approach achieved an area under the curve of 0.934 for men and 0.932 for women, outperforming the original Swedish-based EXINIbone system.
The study utilized BONENAVI version 2, which was trained on 1,532 scans from nine different institutions. This tool calculates the probability of abnormality and the bone scan index to assist clinicians in evaluating potential metastatic disease.
The authors state that a large, multi-institutional database is necessary to account for variations in scan quality and patient demographics. This diversity ensures the system remains accurate across different cancer types, such as breast and prostate, rather than relying on a single-institution or single-region dataset.
The researchers used a validation set of 503 multi-center bone scans. This dataset included 207 patients with prostate cancer and 166 patients with breast cancer, allowing for a robust assessment of the software's performance across different clinical populations.
The study measured diagnostic performance using receiver operating characteristic analysis and net reclassification improvement. These metrics allowed the researchers to compare the sensitivity and specificity of the different software versions, with the final version achieving 90% sensitivity and 84% specificity in men.
The authors claim that their findings demonstrate the importance of using a sufficient number of diverse training databases. They suggest that future diagnostic systems should prioritize large-scale, multi-center data collection to ensure high performance across varied patient populations.

![Quantitative [18F]-Naf-PET-MRI Analysis for the Evaluation of Dynamic Bone Turnover in a Patient with Facetogenic Low Back Pain](/_next/image?url=https%3A%2F%2Fcloudfront.jove.com%2FCDNSource%2Fteasers%2F58491.jpg&w=3840&q=50)