The Utility of a Deep Learning-based Algorithm for Bone Scintigraphy in Patient with Prostate Cancer

Objective: Bone scintigraphy has often been used to evaluate bone metastases. Its functionality is evident in detecting bone metastasis in patients with malignant tumor including prostate cancer, as appropriate treatment and prognosis are dependent on the presence and degree of bone metastasis. The development of a deep learning-based algorithm in the field of information processing has been remarkable in recent years. We hypothesized that a deep learning-based algorithm is useful in diagnosing osseous metastases in patients with prostate cancer using bone scintigraphy. Thus, this study aims to examine the utility of deep learning-based algorithm in detecting bone metastases in patients with prostate cancer, as compared with nuclear medicine specialists.

Methods: In total, 139 serial patients with prostate cancer, who underwent whole-body bone scintigraphy, were enrolled in this study. Each scintigraphy examination was evaluated visually and independently by nuclear medicine specialists; this was also analyzed using a deep learning-based algorithm. The number of abnormal uptakes was assessed by the nuclear medicine specialists and with a software which used the deep learning-based algorithm, and the per-patient detection rate and the per-region detection rate were then calculated. The software automatically analyzed bone scintigraphy for the presence or absence of osseous metastasis in individual patients, for the 12 body regions. The detection rates analyzed separately by the nuclear medicine specialists and using the software were then compared. The sensitivity, specificity, and accuracy by the specialist and with the software were calculated.

Results: The sensitivity, specificity, and accuracy by the nuclear medicine specialists were 100%, 94.9% and 97.1%. On the other hand, they with the software were 91.7%, 87.3% and 89.2%. No statistically significant difference was determined between the per-patient detection rates assessed by the specialists versus the software. In regional assessment, there was also no statistically significant difference between most of the per-region detection rates (10 of 12 regions) by the specialists versus the results obtained by the software.

Conclusions: The software with the deep learning-based algorithm might be used as diagnostic aid in the evaluation of bone metastases for prostate cancer patients.